diff --git a/.codex b/.codex
deleted file mode 100644
index e69de29..0000000
diff --git a/.github/ISSUE_TEMPLATE/bug_report.yml b/.github/ISSUE_TEMPLATE/bug_report.yml
new file mode 100644
index 0000000..89ae507
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/bug_report.yml
@@ -0,0 +1,84 @@
+name: Bug report
+description: Something in xrag behaves incorrectly
+labels: ["needs-triage"]
+body:
+  - type: markdown
+    attributes:
+      value: |
+        Thanks for filing a bug. Please redact all API keys and any
+        confidential document content before submitting.
+
+        For security vulnerabilities, follow
+        [SECURITY.md](https://github.com/henryle97/xrag/blob/main/SECURITY.md)
+        instead of opening a public issue.
+
+  - type: textarea
+    id: what-happened
+    attributes:
+      label: What happened?
+      description: What did you observe, and what did you expect instead?
+    validations:
+      required: true
+
+  - type: textarea
+    id: reproduce
+    attributes:
+      label: Steps to reproduce
+      description: >
+        A minimal snippet or CLI invocation. Include the relevant config with
+        credentials redacted.
+      render: shell
+    validations:
+      required: true
+
+  - type: input
+    id: version
+    attributes:
+      label: xrag version
+      description: "Output of `python -c \"import xrag; print(xrag.__version__)\"`"
+      placeholder: "0.5.3"
+    validations:
+      required: true
+
+  - type: input
+    id: python-version
+    attributes:
+      label: Python version
+      placeholder: "3.12.9"
+    validations:
+      required: true
+
+  - type: dropdown
+    id: area
+    attributes:
+      label: Which area?
+      options:
+        - parser
+        - chunker
+        - enrichment
+        - embedding
+        - indexing / vector store
+        - retrieval / rerank
+        - generation
+        - Xrag client
+        - CLI
+        - eval / benchmark
+        - packaging / install
+        - not sure
+    validations:
+      required: true
+
+  - type: textarea
+    id: logs
+    attributes:
+      label: Logs or traceback
+      description: Paste the full traceback if there is one. Redact credentials.
+      render: shell
+
+  - type: textarea
+    id: context
+    attributes:
+      label: Anything else?
+      description: >
+        Vector store backend, provider models, document type, or anything else
+        that helps reproduce.
diff --git a/.github/ISSUE_TEMPLATE/config.yml b/.github/ISSUE_TEMPLATE/config.yml
new file mode 100644
index 0000000..c91fd72
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/config.yml
@@ -0,0 +1,8 @@
+blank_issues_enabled: false
+contact_links:
+  - name: Question or usage help
+    url: https://github.com/henryle97/xrag/discussions
+    about: Ask how to use xrag, or discuss an idea before filing a feature request.
+  - name: Security vulnerability
+    url: https://github.com/henryle97/xrag/security/advisories/new
+    about: Report privately. Please do not open a public issue for security problems.
diff --git a/.github/ISSUE_TEMPLATE/feature_request.yml b/.github/ISSUE_TEMPLATE/feature_request.yml
new file mode 100644
index 0000000..dfbddc5
--- /dev/null
+++ b/.github/ISSUE_TEMPLATE/feature_request.yml
@@ -0,0 +1,54 @@
+name: Feature request
+description: Suggest a capability or improvement for xrag
+labels: ["needs-triage"]
+body:
+  - type: textarea
+    id: problem
+    attributes:
+      label: What problem are you trying to solve?
+      description: >
+        Describe the use case rather than a specific implementation. What are
+        you unable to do today?
+    validations:
+      required: true
+
+  - type: textarea
+    id: proposal
+    attributes:
+      label: Proposed solution
+      description: What would the API, config, or CLI surface look like?
+    validations:
+      required: true
+
+  - type: textarea
+    id: alternatives
+    attributes:
+      label: Alternatives considered
+      description: Workarounds you tried, or other designs you weighed.
+
+  - type: dropdown
+    id: area
+    attributes:
+      label: Which area?
+      options:
+        - parser
+        - chunker
+        - enrichment
+        - embedding
+        - indexing / vector store
+        - retrieval / rerank
+        - generation
+        - Xrag client
+        - CLI
+        - eval / benchmark
+        - packaging / install
+        - not sure
+    validations:
+      required: true
+
+  - type: checkboxes
+    id: contribution
+    attributes:
+      label: Contribution
+      options:
+        - label: I'd be willing to open a PR for this
diff --git a/.github/dependabot.yml b/.github/dependabot.yml
new file mode 100644
index 0000000..e9bc56f
--- /dev/null
+++ b/.github/dependabot.yml
@@ -0,0 +1,36 @@
+version: 2
+updates:
+  # Keep CI actions current — these are the highest-value updates for a
+  # library, since a compromised action runs with repo credentials.
+  - package-ecosystem: github-actions
+    directory: "/"
+    schedule:
+      interval: weekly
+    labels: ["needs-triage"]
+    commit-message:
+      prefix: "ci"
+
+  # Python dependencies. Grouped so routine patch bumps arrive as one PR
+  # rather than flooding the queue.
+  - package-ecosystem: uv
+    directory: "/"
+    schedule:
+      interval: weekly
+    open-pull-requests-limit: 5
+    labels: ["needs-triage"]
+    commit-message:
+      prefix: "chore(deps)"
+    groups:
+      patch-and-minor:
+        update-types:
+          - minor
+          - patch
+    ignore:
+      # LangChain majors are pinned deliberately and exercised by the contract
+      # matrix in ci.yml. Bump these by hand alongside that matrix.
+      - dependency-name: "langchain"
+        update-types: ["version-update:semver-major"]
+      - dependency-name: "langchain-core"
+        update-types: ["version-update:semver-major"]
+      - dependency-name: "langchain-openai"
+        update-types: ["version-update:semver-major"]
diff --git a/.github/pull_request_template.md b/.github/pull_request_template.md
index ce1fed4..9faf2a8 100644
--- a/.github/pull_request_template.md
+++ b/.github/pull_request_template.md
@@ -1,5 +1,5 @@
 <!--
-xrag PR template. See docs/contributing.md §3 for the conventions
+xrag PR template. See CONTRIBUTING.md §3 for the conventions
 this template encodes. Keep the four-bullet Summary shape; skip the
 Risks section entirely when there are none rather than writing "None".
 -->
@@ -24,7 +24,7 @@ ONE `Closes #N` per issue, each on its own content line.
 GitHub's auto-close parser only fires on `Closes`/`Fixes`/`Resolves`
 when the keyword is on the same content line as the reference — a
 heading like `## Closes` followed by `#19, #20, #21` will silently
-fail to close anything. See docs/contributing.md §3.1.
+fail to close anything. See CONTRIBUTING.md §3.1.
 
 Single-issue example:
   Closes #42
diff --git a/.gitleaksignore b/.gitleaksignore
new file mode 100644
index 0000000..8ef8dd8
--- /dev/null
+++ b/.gitleaksignore
@@ -0,0 +1,17 @@
+# Reviewed gitleaks false positives.
+#
+# Fingerprints are <path>:<rule>:<line> and are line-sensitive: if one of
+# these lines moves, gitleaks re-reports it and the fingerprint below needs
+# updating. Re-verify that the match is still a placeholder before doing so —
+# never add a fingerprint without reading the line it points at.
+
+# `dedup_family=True` is a retrieval config flag. The generic-api-key rule
+# matches the `<word>=<value>` shape, not anything secret. This entry is a
+# table of doc snippets asserted to stay in sync with the default config.
+tests/unit/test_client_app_config.py:generic-api-key:94
+
+# `Authorization: Bearer rk_live_...` in an API design document. The literal
+# trailing ellipsis is part of the text — this is an illustrative curl example
+# for a proposed per-tenant key format, not an issued credential.
+docs/sub-plans/api/apis_design.md:curl-auth-header:685
+docs/sub-plans/api/apis_design.md:curl-auth-header:705
diff --git a/.pre-commit-config.yaml b/.pre-commit-config.yaml
index 8ecfb74..4540952 100644
--- a/.pre-commit-config.yaml
+++ b/.pre-commit-config.yaml
@@ -42,6 +42,9 @@ repos:
         args: [--branch, main]
 
   - repo: https://github.com/gitleaks/gitleaks
-    rev: v8.21.2
+    # Keep this reasonably current. Binaries built from older revs (v8.21.2
+    # and earlier) fail to load on recent macOS with a dyld "missing LC_UUID"
+    # abort, which surfaces as a hook crash (exit -6) rather than a finding.
+    rev: v8.30.1
     hooks:
       - id: gitleaks
diff --git a/CHANGELOG.md b/CHANGELOG.md
new file mode 100644
index 0000000..02a141d
--- /dev/null
+++ b/CHANGELOG.md
@@ -0,0 +1,106 @@
+# Changelog
+
+All notable changes to this project are documented here.
+
+The format is based on [Keep a Changelog](https://keepachangelog.com/en/1.1.0/),
+and this project adheres to [Semantic Versioning](https://semver.org/spec/v2.0.0.html).
+
+Only names exported from `xrag.__all__` are covered by the version contract.
+Anything outside that surface is internal and may change in any release.
+
+## [Unreleased]
+
+### Added
+
+- DOCX image extraction: `DocxPicturePartitioner` and an `extract_images`
+  parser option (#61)
+- Image chunks in `SectionTableChunker` (#62)
+- LLM-generated image descriptions (#64)
+
+### Changed
+
+- Aligned retrieval and indexing defaults (#67)
+
+## [0.5.3] — 2026-05-05
+
+### Added
+
+- Navigational testset generator, with ADR-0003 recording the primary-chunk
+  evidence decision (#51)
+- Navigation questions in dataset generation (#54)
+- Per-category metric breakdown by `reasoning_type` (#52)
+- Append mode for the `eval create-dataset` CLI (#53)
+
+### Changed
+
+- Indexing now augments retrieval text instead of replacing it, per ADR-0002
+  (#45)
+
+## [0.5.2] — 2026-05-05
+
+### Fixed
+
+- Pin `openai_api_type=None` in the embedder to stop an environment variable
+  leaking into requests (#42)
+
+## [0.5.1] — 2026-05-05
+
+### Added
+
+- `ChunkMetadata` promoted to the public `__all__` surface (#40)
+
+## [0.5.0] — 2026-05-04
+
+### Added
+
+- Per-step `max_concurrency` for LLM-bound enrichers (#39)
+
+### Changed
+
+- **Breaking:** unified the chunk schema on the canonical V2 envelope and
+  `doc_id` (#37)
+
+## [0.4.1] — 2026-05-04
+
+### Added
+
+- `qdrant_api_key` parameter on `Xrag.__init__` (#29)
+
+## [0.4.0] — 2026-05-03
+
+### Changed
+
+- Explicit settings injection across the asymmetric boundary, recorded in
+  ADR-0001 (#26)
+
+## [0.3.0] — 2026-05-01
+
+### Changed
+
+- Library and client SDK cleanup pass across the public surface (#3)
+
+## [0.2.0] — 2026-04-27
+
+### Added
+
+- High-level async `Xrag` client with resource-namespaced operations
+  (`documents.*`, `retrievals.*`, `rag.*`)
+- Tenant binding via `client.for_tenant(...)`
+- Typed error hierarchy rooted at `XragError`
+- Split the local-PDF parser into its own `parser-unstructured-local-pdf` extra
+
+### Changed
+
+- Packaging moved to the hatch backend; the project publishes to PyPI as
+  `pyxrag` while the import name stays `xrag`
+- Renamed the engine repository from `fin-rag` to `xrag`
+
+[Unreleased]: https://github.com/henryle97/xrag/compare/v0.5.3...HEAD
+[0.5.3]: https://github.com/henryle97/xrag/compare/v0.5.2...v0.5.3
+[0.5.2]: https://github.com/henryle97/xrag/compare/v0.5.1...v0.5.2
+[0.5.1]: https://github.com/henryle97/xrag/compare/v0.5.0...v0.5.1
+[0.5.0]: https://github.com/henryle97/xrag/compare/v0.4.1...v0.5.0
+[0.4.1]: https://github.com/henryle97/xrag/compare/v0.4.0...v0.4.1
+[0.4.0]: https://github.com/henryle97/xrag/compare/v0.3.0...v0.4.0
+[0.3.0]: https://github.com/henryle97/xrag/compare/v0.2.0...v0.3.0
+[0.2.0]: https://github.com/henryle97/xrag/releases/tag/v0.2.0
diff --git a/CLAUDE.md b/CLAUDE.md
index 0221fe0..0cd76ac 100644
--- a/CLAUDE.md
+++ b/CLAUDE.md
@@ -83,5 +83,3 @@ Single-context repo — `CONTEXT.md` and `docs/adr/` at the repo root (created l
 ## Rules
 
 - gh use the `henryle97` account. Run `gh auth switch -u henryle97` if needed
-
-
diff --git a/CODE_OF_CONDUCT.md b/CODE_OF_CONDUCT.md
new file mode 100644
index 0000000..6bb9929
--- /dev/null
+++ b/CODE_OF_CONDUCT.md
@@ -0,0 +1,126 @@
+# Contributor Covenant Code of Conduct
+
+## Our Pledge
+
+We as members, contributors, and leaders pledge to make participation in our
+community a harassment-free experience for everyone, regardless of age, body
+size, visible or invisible disability, ethnicity, sex characteristics, gender
+identity and expression, level of experience, education, socio-economic status,
+nationality, personal appearance, race, caste, color, religion, or sexual
+identity and orientation.
+
+We pledge to act and interact in ways that contribute to an open, welcoming,
+diverse, inclusive, and healthy community.
+
+## Our Standards
+
+Examples of behavior that contributes to a positive environment:
+
+- Demonstrating empathy and kindness toward other people
+- Being respectful of differing opinions, viewpoints, and experiences
+- Giving and gracefully accepting constructive feedback
+- Accepting responsibility and apologizing to those affected by our mistakes,
+  and learning from the experience
+- Focusing on what is best not just for us as individuals, but for the overall
+  community
+
+Examples of unacceptable behavior:
+
+- The use of sexualized language or imagery, and sexual attention or advances of
+  any kind
+- Trolling, insulting or derogatory comments, and personal or political attacks
+- Public or private harassment
+- Publishing others' private information, such as a physical or email address,
+  without their explicit permission
+- Other conduct which could reasonably be considered inappropriate in a
+  professional setting
+
+## Enforcement Responsibilities
+
+Community leaders are responsible for clarifying and enforcing our standards of
+acceptable behavior and will take appropriate and fair corrective action in
+response to any behavior that they deem inappropriate, threatening, offensive,
+or harmful.
+
+Community leaders have the right and responsibility to remove, edit, or reject
+comments, commits, code, wiki edits, issues, and other contributions that are
+not aligned to this Code of Conduct, and will communicate reasons for moderation
+decisions when appropriate.
+
+## Scope
+
+This Code of Conduct applies within all community spaces, and also applies when
+an individual is officially representing the community in public spaces.
+Examples include using an official email address, posting via an official social
+media account, or acting as an appointed representative at an online or offline
+event.
+
+## Enforcement
+
+Instances of abusive, harassing, or otherwise unacceptable behavior may be
+reported to the project maintainers by opening a
+[private security advisory](https://github.com/henryle97/xrag/security/advisories/new),
+which is visible only to maintainers, or by contacting
+[@henryle97](https://github.com/henryle97) directly on GitHub.
+
+All complaints will be reviewed and investigated promptly and fairly. All
+community leaders are obligated to respect the privacy and security of the
+reporter of any incident.
+
+## Enforcement Guidelines
+
+Community leaders will follow these Community Impact Guidelines in determining
+the consequences for any action they deem in violation of this Code of Conduct:
+
+### 1. Correction
+
+**Community Impact**: Use of inappropriate language or other behavior deemed
+unprofessional or unwelcome in the community.
+
+**Consequence**: A private, written warning from community leaders, providing
+clarity around the nature of the violation and an explanation of why the
+behavior was inappropriate. A public apology may be requested.
+
+### 2. Warning
+
+**Community Impact**: A violation through a single incident or series of
+actions.
+
+**Consequence**: A warning with consequences for continued behavior. No
+interaction with the people involved, including unsolicited interaction with
+those enforcing the Code of Conduct, for a specified period of time. Violating
+these terms may lead to a temporary or permanent ban.
+
+### 3. Temporary Ban
+
+**Community Impact**: A serious violation of community standards, including
+sustained inappropriate behavior.
+
+**Consequence**: A temporary ban from any sort of interaction or public
+communication with the community for a specified period of time. Violating these
+terms may lead to a permanent ban.
+
+### 4. Permanent Ban
+
+**Community Impact**: Demonstrating a pattern of violation of community
+standards, including sustained inappropriate behavior, harassment of an
+individual, or aggression toward or disparagement of classes of individuals.
+
+**Consequence**: A permanent ban from any sort of public interaction within the
+community.
+
+## Attribution
+
+This Code of Conduct is adapted from the [Contributor Covenant][homepage],
+version 2.1, available at
+https://www.contributor-covenant.org/version/2/1/code_of_conduct.html.
+
+Community Impact Guidelines were inspired by
+[Mozilla's code of conduct enforcement ladder][mozilla].
+
+For answers to common questions about this code of conduct, see the FAQ at
+https://www.contributor-covenant.org/faq. Translations are available at
+https://www.contributor-covenant.org/translations.
+
+[homepage]: https://www.contributor-covenant.org
+[mozilla]: https://github.com/mozilla/diversity
diff --git a/CONTRIBUTING.md b/CONTRIBUTING.md
new file mode 100644
index 0000000..484dcbb
--- /dev/null
+++ b/CONTRIBUTING.md
@@ -0,0 +1,150 @@
+# Contributing to xrag
+
+Thanks for your interest in xrag. This document covers how to get set up, the
+checks we expect before review, and the conventions the PR template encodes.
+
+## 1. Getting set up
+
+xrag uses [uv](https://docs.astral.sh/uv/) for dependency management and
+requires Python 3.11 or newer.
+
+```bash
+git clone https://github.com/henryle97/xrag.git
+cd xrag
+make sync              # uv sync --group dev + the extras the tests need
+make precommit-install # installs the pre-commit hooks (recommended)
+```
+
+`make sync` installs the `cli`, `chroma`, and `rerank-cohere` extras because
+the unit tests import them directly. A bare `uv sync` is not enough to run the
+test suite.
+
+Local PDF/OCR parsing is heavy and optional. Add it only if you're working on
+that path:
+
+```bash
+uv sync --group dev --extra cli --extra chroma --extra rerank-cohere \
+  --extra parser-unstructured-local-pdf
+```
+
+### Environment variables
+
+Copy `.env.example` to `.env` and fill in only the keys you need. Every
+variable maps to a field on `xrag.config.settings.Settings`. Unit tests are
+hermetic and need no credentials.
+
+## 2. Development workflow
+
+Run these before opening a PR:
+
+```bash
+make lint        # ruff check + ruff format --check
+make unit-test   # pytest tests/unit -q — hermetic, this is the CI gate
+```
+
+`make check` additionally runs `compileall`. Integration tests
+(`make integration-test`) make real network calls, are not part of CI, and
+require credentials.
+
+After changing runtime behavior, also exercise the changed path with a live CLI
+command — unit tests alone don't prove a pipeline still runs end to end.
+
+### Tests
+
+- Unit tests live in `tests/unit/` and must stay hermetic — no network, no
+  credentials, no reliance on files outside the repo.
+- Contract tests in `tests/contract/` run against the built wheel and guard the
+  public API surface. If you add or remove anything in `xrag.__all__`, update
+  them.
+- Mark network-dependent tests with `requires_network`, `requires_qdrant`, or
+  `requires_openai`.
+
+### Public API surface
+
+Everything exported from `xrag.__all__` is the supported surface and is
+SemVer-bound. Anything outside it is internal. Changing an exported name is a
+breaking change and needs a contract-test update plus a changelog entry.
+
+## 3. Pull requests
+
+The [PR template](.github/pull_request_template.md) asks for four things.
+Keep the shape:
+
+- **Problem** — what's broken or missing
+- **Why** — why it's worth changing now
+- **What changed** — the actual diff, in prose
+- **What did NOT change** — scope boundaries, so reviewers know where not to look
+
+Include a **Test plan** with `make lint` and `make unit-test` checked, plus any
+live CLI, integration, or manual verification you did. Skip the Risks section
+entirely when there are none rather than writing "None".
+
+Other expectations:
+
+- Branch from an up-to-date `main`: `git checkout main && git pull --ff-only origin main`
+- Keep PRs focused. Parent/child issues ship together as the parent's PR.
+- Resolve review threads only after the fix is pushed.
+- Don't auto-merge.
+
+### 3.1 Closing issues
+
+GitHub's auto-close parser only fires when a close keyword (`Closes`, `Fixes`,
+`Resolves`) sits on the same content line as the issue reference. A heading like
+`## Closes` followed by `#19, #20, #21` **silently closes nothing**.
+
+Write one close keyword per issue, each on its own content line:
+
+```text
+Closes #19. Closes #20. Closes #21.
+```
+
+Verify before merging:
+
+```bash
+gh pr view <N> --json closingIssuesReferences
+```
+
+## 4. Commit conventions
+
+We use [Conventional Commits](https://www.conventionalcommits.org/) with a
+module scope:
+
+```text
+feat(parser): add DocxPicturePartitioner
+fix(config): reject unknown provider names
+test(chunker): cover image chunk emission
+docs(plans): refresh indexing backlog
+```
+
+Common scopes: `parser`, `chunker`, `embedder`, `indexer`, `client`, `config`,
+`eval`, `ci`, `packaging`.
+
+## 5. Architecture decisions
+
+Check [`docs/adr/`](docs/adr/) before making architectural changes — it records
+decisions that are easy to accidentally reverse. Add a new ADR when you change
+a cross-module contract or reverse an existing decision.
+
+Deeper domain context lives in [`CONTEXT.md`](CONTEXT.md) and
+[`docs/agents/domain.md`](docs/agents/domain.md).
+
+## 6. Evaluation and benchmarks
+
+Retrieval quality changes need evidence. The full discipline is documented in
+[`CLAUDE.md`](CLAUDE.md); the short version:
+
+- Run evals through `scripts/eval/`, which writes to `experiments/eval/<config-stem>-<timestamp>/`.
+- **Append** a new row to [`docs/BENCHMARK.md`](docs/BENCHMARK.md) — never edit
+  existing rows.
+- Per-category `R@10` is the primary regression signal.
+- `experiments/eval/experiments.jsonl` is the append-only source of truth.
+
+Benchmark documentation uses a fictional issuer and synthetic figures. Never
+commit real customer documents, real financial data, or anything under `data/`
+(it is gitignored for this reason).
+
+## 7. Reporting bugs and requesting features
+
+Use the [issue templates](https://github.com/henryle97/xrag/issues/new/choose).
+For security issues, follow [SECURITY.md](SECURITY.md) instead of opening a
+public issue.
diff --git a/Makefile b/Makefile
index 404bc19..bca996c 100644
--- a/Makefile
+++ b/Makefile
@@ -1,7 +1,11 @@
 .PHONY: sync lint format check unit-test integration-test parser parser-smoke precommit-install precommit-run
 
+# Extras the unit tests import directly (chromadb, cohere, typer). Kept in sync
+# with the `unit-test` job in .github/workflows/ci.yml.
+DEV_EXTRAS := --extra cli --extra chroma --extra rerank-cohere
+
 sync:
-	uv sync
+	uv sync --group dev $(DEV_EXTRAS)
 
 lint:
 	uv run ruff check .
@@ -14,7 +18,7 @@ check: lint
 	python3 -m compileall xrag examples tools
 
 unit-test:
-	uv run pytest tests/unit -q
+	uv run $(DEV_EXTRAS) pytest tests/unit -q
 
 integration-test:
 	uv run pytest tests/integration -q
diff --git a/README.md b/README.md
index ee6433f..e7116fa 100644
--- a/README.md
+++ b/README.md
@@ -349,7 +349,7 @@ Recommended path:
 7. Use [`docs/sub-plans/xrag-public-api.md`](docs/sub-plans/xrag-public-api.md) for the high-level Python client design.
 8. Use [`docs/plans.md`](docs/plans.md) and [`docs/sub-plans/`](docs/sub-plans/) for roadmap and implementation details.
 
-## Quickstart — library (v0.1, functional API)
+## Quickstart — library (functional API)
 
 ```python
 from pathlib import Path
@@ -369,7 +369,11 @@ result = run_rag("What was Cash as of December 31, 2024?", cfg)
 print(result.answer)
 ```
 
-The high-level `Xrag` async client (`client.documents.ingest(...)`, `client.retrievals.search(...)`, `client.rag.ask(...)`) ships in v0.2 — see [`docs/sub-plans/xrag-public-api.md`](docs/sub-plans/xrag-public-api.md).
+Most applications should prefer the high-level `Xrag` client shown above. Use
+the functional API when you already work in terms of `AppConfig` and want
+direct control over pipeline wiring — see
+[`docs/sub-plans/xrag-public-api.md`](docs/sub-plans/xrag-public-api.md) for the
+design behind both surfaces.
 
 ## Quickstart — CLI
 
@@ -414,7 +418,7 @@ uv run python -m xrag.cli parser preprocess \
 # Chunk
 uv run python -m xrag.cli chunk prepare \
   --input data/dev/ir_document/normalized/elements.json \
-  --config configs/chunker.yml \
+  --config configs/chunkers/section-table.yml \
   --output-dir data/dev/ir_document/chunked
 
 # Baseline RAG
@@ -447,15 +451,25 @@ UNSTRUCTURED_API_URL=...   # optional
 
 For DOCX-to-PDF conversion install LibreOffice so `soffice` / `libreoffice` is on `PATH`.
 
-## Phasing
+## Project status
 
-| Version | Surface |
-|---|---|
-| **v0.1** (current) | Functional API: `run_ingest`, `run_rag`, Pydantic models, configs, loaders. |
-| **v0.2** | High-level `Xrag` async client with resource-namespaced ops (`documents.*`, `retrievals.*`, `rag.*`). Tenant binding via `for_tenant`. Typed error hierarchy. |
-| **v0.3** | `XragSync` mirror. |
-| **v0.4** | Per-call `config_overrides`, batch ingest, persistent registry contract. |
-| **v0.5+ (backlog)** | Streaming surfaces (`documents.ingest_stream`, `rag.ask_stream`). |
+`xrag` is pre-1.0. Everything exported from `xrag.__all__` is the supported
+surface; anything outside it is internal and not SemVer-bound. Minor releases
+may still change internals.
+
+**Available today:**
+
+- Functional API — `run_ingest`, `run_rag`, Pydantic models, config loaders
+- High-level async `Xrag` client with resource-namespaced ops (`documents.*`,
+  `retrievals.*`, `rag.*`)
+- Tenant binding via `client.for_tenant(...)`
+- Typed error hierarchy (`XragError` and subclasses)
+- Per-call ingest and retrieval overrides
+
+**Planned:**
+
+- `XragSync` — a synchronous mirror of the async client
+- Streaming surfaces (`documents.ingest_stream`, `rag.ask_stream`)
 
 Full design: [`docs/sub-plans/xrag-public-api.md`](docs/sub-plans/xrag-public-api.md).
 
@@ -492,8 +506,8 @@ uv sync --group dev --extra cli --extra chroma --extra rerank-cohere --extra par
 
 - Parser provider: `unstructured`.
 - Preprocess config: [`configs/preprocess.yml`](configs/preprocess.yml).
-- Default chunker: `section_table` in [`configs/chunker.yml`](configs/chunker.yml).
-- Alternative chunker: `ragflow` in [`configs/chunker-ragflow.yml`](configs/chunker-ragflow.yml).
+- Default chunker: `section_table` in [`configs/chunkers/section-table.yml`](configs/chunkers/section-table.yml).
+- Alternative chunkers: `section_token` (alias `ragflow`) in [`configs/chunkers/section-token.yml`](configs/chunkers/section-token.yml), and `title_hierarchy` in [`configs/chunkers/title-hierarchy.yml`](configs/chunkers/title-hierarchy.yml).
 - Simple one-shot RAG config: [`configs/rag-simple-baseline.yml`](configs/rag-simple-baseline.yml).
 - Persistent index/query config: [`configs/rag-baseline.yml`](configs/rag-baseline.yml).
 - Enrichment scenario configs: [`configs/enrichment/`](configs/enrichment).
@@ -507,7 +521,8 @@ uv sync --group dev --extra cli --extra chroma --extra rerank-cohere --extra par
 - [`tools/`](tools/): **dev-only** — eval scoring, dataset generation, HTML viewers, eval datasets. Excluded from the wheel.
 - [`configs/`](configs/): runtime YAML configs.
 - [`docs/`](docs/README.md): user docs, plans, surveys, and terminology.
-- [`data/`](data/): local inputs and generated artifacts.
+- `data/`: local inputs and generated artifacts. Gitignored — created on first
+  run, never committed.
 
 ## Development checks
 
@@ -530,6 +545,17 @@ GitHub Actions now verifies the same core paths contributors should run locally:
 
 Tag pushes matching `v*` reuse the verified package artifacts and publish them to the GitHub Release instead of rebuilding a second time during release.
 
+## Contributing
+
+Contributions are welcome. See [CONTRIBUTING.md](CONTRIBUTING.md) for setup,
+the checks expected before review, and PR conventions. Please also read the
+[Code of Conduct](CODE_OF_CONDUCT.md).
+
+Release history is in [CHANGELOG.md](CHANGELOG.md).
+
+To report a security issue, follow [SECURITY.md](SECURITY.md) rather than
+opening a public issue.
+
 ## License
 
 MIT — see [LICENSE](LICENSE).
diff --git a/SECURITY.md b/SECURITY.md
new file mode 100644
index 0000000..2fc217f
--- /dev/null
+++ b/SECURITY.md
@@ -0,0 +1,52 @@
+# Security Policy
+
+## Supported versions
+
+xrag is pre-1.0. Security fixes land on the latest minor release only. Please
+upgrade before reporting an issue against an older version.
+
+| Version | Supported |
+| ------- | --------- |
+| 0.5.x   | ✅        |
+| < 0.5   | ❌        |
+
+## Reporting a vulnerability
+
+**Please do not open a public issue for security problems.**
+
+Report privately through GitHub's
+[private vulnerability reporting](https://github.com/henryle97/xrag/security/advisories/new)
+for this repository. If that is unavailable to you, open a public issue
+containing only a request for a private contact channel — no details.
+
+Please include:
+
+- A description of the issue and its impact
+- Steps to reproduce, or a proof-of-concept
+- Affected version and relevant configuration (redact all credentials)
+
+You can expect an initial response within 7 days and a status update at least
+every 14 days until the issue is resolved.
+
+## Scope
+
+xrag is a library that talks to third-party services (OpenAI, Cohere, Qdrant,
+Unstructured, AWS Bedrock). Vulnerabilities in those services belong to their
+respective vendors. In scope for this project:
+
+- Credential handling and leakage through logs, artifacts, traces, or errors
+- Tenant isolation failures in `client.for_tenant(...)` or collection scoping
+- Path traversal or unsafe file handling in the parser and artifact-store paths
+- Deserialization or injection issues in config, chunk, or dataset loading
+- Dependency vulnerabilities reachable through xrag's default code paths
+
+## Handling credentials
+
+xrag reads credentials from environment variables only — never from config
+YAML. When filing any issue or sharing a config, redact `OPENAI_API_KEY`,
+`COHERE_API_KEY`, `QDRANT_API_KEY`, `UNSTRUCTURED_API_KEY`, AWS credentials,
+and tracing keys.
+
+This repository runs [gitleaks](https://github.com/gitleaks/gitleaks) as a
+pre-commit hook. Install the hooks with `make precommit-install` so secrets are
+caught before they reach a commit.
diff --git a/configs/baseline.yml b/configs/baseline.yml
index 52973eb..b361652 100644
--- a/configs/baseline.yml
+++ b/configs/baseline.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/enrichment/all.yml b/configs/enrichment/all.yml
index db849b9..fbc9566 100644
--- a/configs/enrichment/all.yml
+++ b/configs/enrichment/all.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/enrichment/auto-keywords.yml b/configs/enrichment/auto-keywords.yml
index f3d4a37..7b8186f 100644
--- a/configs/enrichment/auto-keywords.yml
+++ b/configs/enrichment/auto-keywords.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/enrichment/auto-questions.yml b/configs/enrichment/auto-questions.yml
index 85f46c8..de68b0e 100644
--- a/configs/enrichment/auto-questions.yml
+++ b/configs/enrichment/auto-questions.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/enrichment/table-context.yml b/configs/enrichment/table-context.yml
index f65e99f..774d3d9 100644
--- a/configs/enrichment/table-context.yml
+++ b/configs/enrichment/table-context.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/enrichment/table-summary.yml b/configs/enrichment/table-summary.yml
index 68bcb90..3a73dd2 100644
--- a/configs/enrichment/table-summary.yml
+++ b/configs/enrichment/table-summary.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/raw/ir_document/Pj_AVATAR_F_page_260205.docx"
+  - path: "data/raw/ir_document/ir-sample-a.docx"
     doc_id: pj-doc
 
 parser:
diff --git a/configs/parsers/parse-f1-doc-local.yml b/configs/parsers/parse-f1-doc-local.yml
index a866082..4435f25 100644
--- a/configs/parsers/parse-f1-doc-local.yml
+++ b/configs/parsers/parse-f1-doc-local.yml
@@ -4,7 +4,7 @@ pipeline:
   # artifact_dir: data/dev/ir_document/parser_outputs   # override with --output-dir
 
 sources:
-  - path: "data/raw/ir_document/F1-file.docx"
+  - path: "data/raw/ir_document/ir-sample-b.docx"
     doc_id: f1-doc
 
 parser:
diff --git a/configs/parsers/parse-f1-doc.yml b/configs/parsers/parse-f1-doc.yml
index a2b2120..f11f108 100644
--- a/configs/parsers/parse-f1-doc.yml
+++ b/configs/parsers/parse-f1-doc.yml
@@ -4,7 +4,7 @@ pipeline:
   # artifact_dir: data/dev/ir_document/parser_outputs # defauloverwrite by --output-dir
 
 sources:
-  - path: "data/raw/ir_document/F1-file.docx"
+  - path: "data/raw/ir_document/ir-sample-b.docx"
     doc_id: f1-doc
 
 parser:
diff --git a/configs/parsers/parse-pj-doc-local.yml b/configs/parsers/parse-pj-doc-local.yml
index 59b696a..e2ec9ff 100644
--- a/configs/parsers/parse-pj-doc-local.yml
+++ b/configs/parsers/parse-pj-doc-local.yml
@@ -4,7 +4,7 @@ pipeline:
   # artifact_dir: data/dev/ir_document/parser_outputs   # override with --output-dir
 
 sources:
-  - path: "data/raw/ir_document/Pj_AVATAR_F_page_260205.docx"
+  - path: "data/raw/ir_document/ir-sample-a.docx"
     doc_id: pj-doc
 
 parser:
diff --git a/configs/parsers/parse-pj-doc.yml b/configs/parsers/parse-pj-doc.yml
index 5545da4..5a121c3 100644
--- a/configs/parsers/parse-pj-doc.yml
+++ b/configs/parsers/parse-pj-doc.yml
@@ -4,7 +4,7 @@ pipeline:
   # artifact_dir: data/dev/ir_document/parser_outputs # defauloverwrite by --output-dir
 
 sources:
-  - path: "Pj_AVATAR_F_page_260205.docx"
+  - path: "ir-sample-a.docx"
     doc_id: pj-doc
 
 parser:
diff --git a/configs/rag-avatar-eval.yml b/configs/rag-avatar-eval.yml
index 1e9770c..0aa2299 100644
--- a/configs/rag-avatar-eval.yml
+++ b/configs/rag-avatar-eval.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-baseline-ragflow.yml b/configs/rag-ir-baseline-ragflow.yml
index 7378c0d..a75eaf2 100644
--- a/configs/rag-ir-baseline-ragflow.yml
+++ b/configs/rag-ir-baseline-ragflow.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-baseline.yml b/configs/rag-ir-baseline.yml
index 7b5e19b..e53eb10 100644
--- a/configs/rag-ir-baseline.yml
+++ b/configs/rag-ir-baseline.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-hybrid.yml b/configs/rag-ir-hybrid.yml
index e5a286e..bcd7d78 100644
--- a/configs/rag-ir-hybrid.yml
+++ b/configs/rag-ir-hybrid.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-latest.yml b/configs/rag-ir-latest.yml
index e42287c..c620c89 100644
--- a/configs/rag-ir-latest.yml
+++ b/configs/rag-ir-latest.yml
@@ -1,15 +1,15 @@
 pipeline:
   name: rag-ir-all-docs-latest
   mode: full
-  artifact_dir: /data/lucas/projects/rag/ir_v2/data/dev/ir_document
+  artifact_dir: data/dev/ir_document
   tenant_id: dev
   doc_ids:
   - f1-doc
   - pj-doc
 sources:
-- path: /data/lucas/projects/rag/ir_v2/data/dev/ir_document/raw/F1-file.docx
+- path: data/dev/ir_document/raw/ir-sample-b.docx
   doc_id: f1-doc
-- path: /data/lucas/projects/rag/ir_v2/data/dev/ir_document/raw/Pj_AVATAR_F_page_260205.docx
+- path: data/dev/ir_document/raw/ir-sample-a.docx
   doc_id: pj-doc
 parser:
   provider: unstructured_local
@@ -53,7 +53,7 @@ vector_store:
   provider: qdrant
   options:
     collection_name: ir_document_latest
-    path: /data/lucas/projects/rag/ir_v2/data/dev/ir_document/qdrant_db
+    path: data/dev/ir_document/qdrant_db
     distance: cosine
 indexing:
   exclude_parents_from_index: true
diff --git a/configs/rag-ir-mvp1.yml b/configs/rag-ir-mvp1.yml
index 4a8b637..b5a2d73 100644
--- a/configs/rag-ir-mvp1.yml
+++ b/configs/rag-ir-mvp1.yml
@@ -22,7 +22,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/artifacts/ir_document/raw/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/artifacts/ir_document/raw/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-mvp2-rerank.yml b/configs/rag-ir-mvp2-rerank.yml
index 6b8d2d5..4851353 100644
--- a/configs/rag-ir-mvp2-rerank.yml
+++ b/configs/rag-ir-mvp2-rerank.yml
@@ -16,7 +16,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/artifacts/ir_document/raw/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/artifacts/ir_document/raw/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-mvp2.yml b/configs/rag-ir-mvp2.yml
index f0d9325..db07247 100644
--- a/configs/rag-ir-mvp2.yml
+++ b/configs/rag-ir-mvp2.yml
@@ -14,7 +14,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/artifacts/ir_document/raw/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/artifacts/ir_document/raw/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-mvp3-all-docs.yml b/configs/rag-ir-mvp3-all-docs.yml
index 637c6b8..23c46a2 100644
--- a/configs/rag-ir-mvp3-all-docs.yml
+++ b/configs/rag-ir-mvp3-all-docs.yml
@@ -8,9 +8,9 @@ pipeline:
   doc_ids: [f1-doc, pj-doc]
 
 sources:
-  - path: "data/raw/ir_document/F1-file.docx"
+  - path: "data/raw/ir_document/ir-sample-b.docx"
     doc_id: f1-doc
-  - path: "data/raw/ir_document/Pj_AVATAR_F_page_260205.docx"
+  - path: "data/raw/ir_document/ir-sample-a.docx"
     doc_id: pj-doc
 
 parser:
diff --git a/configs/rag-ir-mvp3.yml b/configs/rag-ir-mvp3.yml
index 60cdc13..56681c2 100644
--- a/configs/rag-ir-mvp3.yml
+++ b/configs/rag-ir-mvp3.yml
@@ -17,7 +17,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/artifacts/ir_document/raw/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/artifacts/ir_document/raw/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir-rerank.yml b/configs/rag-ir-rerank.yml
index 4eadf4b..bb8395f 100644
--- a/configs/rag-ir-rerank.yml
+++ b/configs/rag-ir-rerank.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-ir.yml b/configs/rag-ir.yml
index 6c900c6..0a2034e 100644
--- a/configs/rag-ir.yml
+++ b/configs/rag-ir.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/configs/rag-simple-baseline.yml b/configs/rag-simple-baseline.yml
index b20f10d..4512371 100644
--- a/configs/rag-simple-baseline.yml
+++ b/configs/rag-simple-baseline.yml
@@ -4,7 +4,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/docs/BENCHMARK-chunk.md b/docs/BENCHMARK-chunk.md
index f842e63..d36e389 100644
--- a/docs/BENCHMARK-chunk.md
+++ b/docs/BENCHMARK-chunk.md
@@ -113,7 +113,7 @@ All on gpt-4.1-nano for enrichment + generation, `text-embedding-3-small` for em
 1. **`table_summary` did NOT help on this testset.** Clean A/B at the same retrieval stack (rows 04b vs 05) shows `num_match` *regressed* by −0.037 with the summary on, MRR was effectively flat (−0.003), substring tied. Earlier "row 04 vs row 01" delta was *real* but attributable to the hybrid retrieval stack (BM25 + RRF + dedup_family), not the summary.
 2. **Cohere reranker did NOT help on this testset.** Rows 04 vs 04b show `num_match` flat (0.926 = 0.926), MRR +0.002, substring *worse* −0.026 with rerank on. Rerank costs +3651ms p95 retrieval latency for negative quality movement on a 39-q testset.
 3. **The hybrid retrieval stack alone (BM25 + vector + RRF + dedup_family)** is the lever responsible for the gains over the simple-retrieval baseline. Row 05 (no summary, no rerank, just hybrid) lifts num_match from 0.852 to 0.963 vs row 01 — a +0.111 gain attributable purely to the retriever change. Substring also +0.077, MRR +0.051. With ~no latency cost vs simple (450ms vs 437ms p95).
-4. **Likely cause for the summary regression**: hallucinated numbers in summaries (observed in the inspection — chunk:56 "hundreds of millions" vs actual $8K–$438K; chunk:129 "$672M" vs actual ~$672K) shift the embedding into the wrong neighborhood. The LLM at prompt time still reads correct HTML, but it gets the *wrong table* fed in because retrieval ranked elsewhere. `auto_questions` (HyQE) is doing better natural-language anchoring without the numeric hallucination risk because the questions are query-shaped and don't have to commit to specific magnitudes.
+4. **Likely cause for the summary regression**: hallucinated numbers in summaries (observed in the inspection — chunk:56 "hundreds of millions" vs actual $8K–$471K; chunk:129 "$540M" vs actual ~$540K) shift the embedding into the wrong neighborhood. The LLM at prompt time still reads correct HTML, but it gets the *wrong table* fed in because retrieval ranked elsewhere. `auto_questions` (HyQE) is doing better natural-language anchoring without the numeric hallucination risk because the questions are query-shaped and don't have to commit to specific magnitudes.
 5. **`top_k_redundancy` improvements** (0.431 → 0.285–0.295 across rows 04/04b/05) are entirely from `dedup_family: true`, not from any enrichment.
 
 ### Practical takeaways
@@ -161,7 +161,7 @@ All on gpt-4.1-nano for enrichment + generation, `text-embedding-3-small` for em
    string fragment across multiple chunks.
 6. **`exact_match=0.000` everywhere.** The generator's normalized
    answer never matches the gold string verbatim. Expected — `gpt-4.1-nano`
-   paraphrases ("$7,351,900" vs "$7.35M"). Use `numeric_match` /
+   paraphrases ("$6,120,400" vs "$6.12M"). Use `numeric_match` /
    `substring_match` as the answer-grounding signal.
 
 ## Caveats / methodology notes
diff --git a/docs/BENCHMARK.md b/docs/BENCHMARK.md
index 1bb5c16..ba04157 100644
--- a/docs/BENCHMARK.md
+++ b/docs/BENCHMARK.md
@@ -49,7 +49,7 @@ testset:
    from it. **A correct refusal IS the right answer**, so a low score
    here is a metric artifact, not a generation defect.
 2. **Terse factual answers**: byte-exact replies like `"17%"`,
-   `"January 2025"`, `"$4,431,280"` can be the answer to many unrelated
+   `"January 2025"`, `"$3,845,200"` can be the answer to many unrelated
    questions, so the synthetic-question step finds many low-similarity
    candidates and drives the mean down. Several **exact-match** answers
    in the multi-doc 87-Q set (1658 run) score 0.13–0.20.
@@ -165,7 +165,7 @@ count would bring it under 1000ms.
 
 ---
 
-## IR document (AVATAR / Neksai) — 40 LLM-generated questions
+## IR document (AVATAR / Contoso) — 40 LLM-generated questions
 
 Dataset: `data/artifacts/ir_document/eval-llm/testset.jsonl` (40 Qs, 39 valid after validation).
 Categories: table_lookup (8), direct_lookup (15), semantic_paraphrase (8), multi_hop (6), no_answer (3).
@@ -221,7 +221,7 @@ Embedding: `text-embedding-3-small` · Judge: `gpt-4.1-mini` · RAGAS metrics: `
 | 01 | rag-ir-mvp1 | section_table | hybrid | 10 | gpt-4.1-mini | 0.973 | **0.96** | **0.83** | — | 543ms | **6267ms** | 622s | [100q-20260424-1126](../experiments/eval/rag-ir-mvp1-100q-20260424-1126/) |
 | 02 | rag-ir-mvp2 | section_table | hybrid | 10 | gpt-4.1-mini | 0.973 | **0.96** | 0.82 | — | **501ms** | 62791ms§ | 641s | [100q-20260424-1137](../experiments/eval/rag-ir-mvp2-100q-20260424-1137/) |
 
-§ MVP2 `e2e p95=62791ms` is inflated by 5 queries with 60–91s generation times. Spot-check shows correct short answers (e.g. "$22,265,412" in 105 chars) — these are transient `gpt-4.1-mini` API stalls, not prompt-size or retrieval issues. MVP1 (same generator, ran 10 min earlier) saw max gen=9746ms. **Re-run recommended** to confirm MVP2 is not systematically slower.
+§ MVP2 `e2e p95=62791ms` is inflated by 5 queries with 60–91s generation times. Spot-check shows correct short answers (e.g. "$24,830,170" in 105 chars) — these are transient `gpt-4.1-mini` API stalls, not prompt-size or retrieval issues. MVP1 (same generator, ran 10 min earlier) saw max gen=9746ms. **Re-run recommended** to confirm MVP2 is not systematically slower.
 
 ### MVP1 vs MVP2 on 100 Qs
 
@@ -241,7 +241,7 @@ Quality is a statistical tie. The RAGFlow-style indexing changes in MVP2 (parent
 
 ## PJ-doc — 48 LLM-generated questions
 
-New corpus (PJ doc, distinct from AVATAR/Neksai). Indexed into Qdrant collection
+New corpus (PJ doc, distinct from AVATAR/Contoso). Indexed into Qdrant collection
 `ir_dev_mvp3_table_summary_v1` from `data/dev/ir_document/pj-doc/chunked-section-table/chunks.json`
 (174 chunks: 60 section, 24 table, 90 text).
 
@@ -390,7 +390,7 @@ near-perfect on recall.
 | **Faithfulness jumps 0.92 → 0.96 on 100 Qs vs 40 Qs** | Larger testset has a more balanced mix; multi_hop (the 40Q drag, fa=0.72) is proportionally less represented. Not a regression of the smaller testset. |
 | **`gpt-4.1-mini` has sporadic multi-second gen stalls** | MVP2-100q had 5 queries at 60–91s gen (vs MVP1-100q max 9.7s same day). Answers correct and short — provider-side tail, not prompt-size. Inflates e2e p95; mean/median unaffected. |
 | **`gpt-4.1-nano` matches `gpt-4.1-mini` quality at ~1/3 cost** (40 Qs) | Row 07 (mini, MVP2) fa=0.93 ar=0.78 cp=0.81 vs Row 08 (nano, MVP2) fa=0.93 ar=0.81 cp=0.77 — within noise on all RAGAS metrics. Nano tightens e2e p95 from 5556ms → 2243ms (new best in the table). Re-run at 100 Qs recommended before promoting to MVP 1. |
-| **PJ-doc is a harder corpus than AVATAR/Neksai on the same stack** | Same `rag-ir-mvp3-table-summary` config: AVATAR 40-Q (proxy: 39-Q rerank-on row, fa≈0.95 historical) vs PJ-doc 48-Q (row 02: fa=0.884, ar=0.706, cp=0.779). ans_rel drops ~0.10 — PJ-doc questions are tougher to answer concisely, even with element-level retrieval at 100%. Multi-hop (7 Qs) and the wider `needs_review` validation rate (16/48) are the likely driver. |
+| **PJ-doc is a harder corpus than AVATAR/Contoso on the same stack** | Same `rag-ir-mvp3-table-summary` config: AVATAR 40-Q (proxy: 39-Q rerank-on row, fa≈0.95 historical) vs PJ-doc 48-Q (row 02: fa=0.884, ar=0.706, cp=0.779). ans_rel drops ~0.10 — PJ-doc questions are tougher to answer concisely, even with element-level retrieval at 100%. Multi-hop (7 Qs) and the wider `needs_review` validation rate (16/48) are the likely driver. |
 | **`element_*` metrics are the right signal on LLM-gen testsets with overlapping chunks** | PJ-doc row 02: `retrieval_recall=0.761` but `element_recall=1.000`, `element_hit=1.000`, `element_mrr@10=0.893`. The LLM generator labels one chunk-id as evidence, but the retriever often returns a sibling chunk containing the same elements. Treat `element_recall`/`element_mrr` as the truth for retrieval health on this kind of testset. |
 | **Reranker is correctly off by default for inner-loop iteration** (PJ-doc) | Empirically validated by the 39-Q comparison cited in `configs/rag-ir-mvp3-table-summary.yml`: Cohere `rerank-v3.5` adds 6.6× retrieval p95 (649ms → 4300ms) for ~0 quality gain (`element_mrr@10` 0.887 vs 0.889, `substring_match` regressed −0.026). Flip on only for promotion-grade runs. |
 | **Multi-doc element_recall holds at 0.967 (no per-question doc filter)** | Multi-doc row 02 (PJ 50 + F1 49 = 99 Qs) on the combined corpus, both docs visible to the retriever: `element_recall=0.967`, `element_mrr@10=0.724`. Per-doc: PJ 0.880 / F1 0.898. PJ has 46% cross-doc top-10 leakage (smaller corpus, F1 chunks crowd in) and F1 has 14% — both still find chunks carrying the right source elements ~9/10 times. Cross-doc retrieval is *intentional* here, not a fix-it. The earlier 0.494 element_recall (row 01) was a stale-testset artifact: PJ's `evidence_element_ids` were generated against the API parse and didn't survive the local re-parse (0/77 element_ids matched). Regenerating the PJ testset against the local-parsed chunks fixed the metric without any retrieval changes. **Lesson: re-run testset generation any time the parser changes; chunk_ids stay stable but element_ids do not.** |
@@ -400,7 +400,7 @@ near-perfect on recall.
 | **Indexing-time chunk hygiene removes index pollution** (758 vs 778 prior) | New `IndexingConfig` (`min_section_body_chars=30`, `merge_continuation_sections=True`) drops heading-only sections and merges page-split `(cont.)` siblings before parent exclusion: pj-doc 173 → 165 (8 cont. merged) → 113 indexed; f1-doc 840 → 645 indexed (no cont. sections). Cleaner top-k (no embedding budget on `"10. TAXATION"` or cover-page metadata), and `(cont.)` family dedup now treats split sections as one unit. |
 | **`exclude_parents_from_index` moved from env var to YAML** (`AppConfig.indexing`) | `Settings.exclude_parents_from_index` (env var `EXCLUDE_PARENTS_FROM_INDEX`) removed in favor of YAML `indexing.exclude_parents_from_index` (default `True`). Configs are now the single source of truth for "how chunks were indexed"; no per-process env to forget. Two new sibling fields land in the same block: `min_section_body_chars` and `merge_continuation_sections`. |
 | **Enrichment concurrency (max_concurrency=10) is benchmark-neutral** | Multi-doc row 05 (20260504-1511, conc=10) vs row 13 compile log (20260427-0523, conc=1, same 86-Q run): Recall@10 0.937 vs 0.924 (+0.013), MRR@10 0.621 vs 0.641 (−0.020), num_match 0.923 vs 0.897 (+0.026). All deltas within LLM generation variance at temperature=0.1. PR #39 (per-step `batch_as_completed`) correctness confirmed. |
-| **ADR 0002 augment-not-replace is benchmark-neutral on entity-prefixed testsets** (Multi-doc 95 Qs) | Multi-doc row 06 (20260505-1054) vs prior 95-Q best (20260504-1603): R@10 0.994 → 0.983 (−0.011), element_recall 1.000 → 0.989 (−0.011), MRR@10 0.723 → 0.724 (+0.001), num_match 0.750 → 0.750 (flat), retrieval p95 618ms → 433ms (−30%). All deltas within loose tolerance. The one flipped question (`qa:0011`: *"What was the cash balance as of December 31, 2023…"*) is a cell-value table query — exactly one of the two failure modes ADR 0002 targets — and regressed from rank 9 → out-of-top-10 because augmenting bodies into all chunks made f1-doc text chunks more competitive on this query, out-ranking the pj-doc table chunk despite its `Cash \| $4,431,280` now being in the dense vector. **Lesson: this LLM-gen testset (auto-question-derived → entity-prefixed) cannot witness the navigational/cell-value wins ADR 0002 targets; the deferred integration regression pins (PR #45 follow-up) are the right gate.** See `experiments/eval/rag-ir-mvp3-all-docs-20260505-1054/fail_cases.md`. |
+| **ADR 0002 augment-not-replace is benchmark-neutral on entity-prefixed testsets** (Multi-doc 95 Qs) | Multi-doc row 06 (20260505-1054) vs prior 95-Q best (20260504-1603): R@10 0.994 → 0.983 (−0.011), element_recall 1.000 → 0.989 (−0.011), MRR@10 0.723 → 0.724 (+0.001), num_match 0.750 → 0.750 (flat), retrieval p95 618ms → 433ms (−30%). All deltas within loose tolerance. The one flipped question (`qa:0011`: *"What was the cash balance as of December 31, 2023…"*) is a cell-value table query — exactly one of the two failure modes ADR 0002 targets — and regressed from rank 9 → out-of-top-10 because augmenting bodies into all chunks made f1-doc text chunks more competitive on this query, out-ranking the pj-doc table chunk despite its `Cash \| $3,845,200` now being in the dense vector. **Lesson: this LLM-gen testset (auto-question-derived → entity-prefixed) cannot witness the navigational/cell-value wins ADR 0002 targets; the deferred integration regression pins (PR #45 follow-up) are the right gate.** See `experiments/eval/rag-ir-mvp3-all-docs-20260505-1054/fail_cases.md`. |
 | **Latest 124-Q row proves the navigational slice is the missing witness** | Multi-doc row 07 (`2026-05-05 14:54 UTC`, `data/dev/ir_document/latest`): legacy categories stay high on recall — direct_lookup `1.000`, semantic_paraphrase `1.000`, table_lookup `0.958`, multi_hop `0.806` — while `navigational` drops to `0.600` R@10 and `0.221` MRR@10. The L2 gate passes cleanly: `0.600 < median(1.000, 1.000, 0.958, 0.806) = 0.979`. Concrete miss cluster: all three `Our Strategy` navigational rows miss their pinned chunk entirely, with `SPECIAL NOTE REGARDING FORWARD-LOOKING STATEMENTS` and TOC-like sections crowding the top ranks instead. See `experiments/eval/rag-ir-mvp3-all-docs-20260505-1454/fail_cases.md`. |
 
 ---
diff --git a/docs/adr/0002-retrieval-text-augment-not-replace.md b/docs/adr/0002-retrieval-text-augment-not-replace.md
index 0c2d2b6..5c7d7d8 100644
--- a/docs/adr/0002-retrieval-text-augment-not-replace.md
+++ b/docs/adr/0002-retrieval-text-augment-not-replace.md
@@ -31,10 +31,10 @@ both query and document.
 Our corpus is F-1 / project-document filings. Users ask **navigational**
 questions like *"how is the corporate structured?"* without naming the
 issuer. Auto-generated questions all start with the entity name (*"What
-strategic corporate restructuring steps did Neksai undertake…"*), so
+strategic corporate restructuring steps did Contoso undertake…"*), so
 entity-less queries miss the dense vector entirely. Issue #43 documents
 the failure case: `f1-doc:chunk:84` (the actual structural narrative
-naming Neksai Inc., Neksai BVI, Neksai SG) never appears in top-10
+naming Contoso Inc., Contoso BVI, Contoso SG) never appears in top-10
 for the query *"how is the corporate structured"*, but does for the
 entity-injected variant *"how is nexksai corporate structured"*.
 
diff --git a/docs/adr/0003-navigational-testset-primary-chunk-evidence.md b/docs/adr/0003-navigational-testset-primary-chunk-evidence.md
index 3031126..2a35899 100644
--- a/docs/adr/0003-navigational-testset-primary-chunk-evidence.md
+++ b/docs/adr/0003-navigational-testset-primary-chunk-evidence.md
@@ -19,8 +19,8 @@ shares: **multiple chunks legitimately satisfy the question**. For
 *"how is the corporate structured?"* against the f1-doc corpus, the
 following chunks are all valid retrieval results:
 
-- `f1-doc:chunk:84` — leaf naming the subsidiaries (*Neksai Inc.*,
-  *Neksai BVI*, *Neksai SG*)
+- `f1-doc:chunk:84` — leaf naming the subsidiaries (*Contoso Inc.*,
+  *Contoso BVI*, *Contoso SG*)
 - the parent section under *"CORPORATE HISTORY AND STRUCTURE"* —
   describes the structure narratively
 - a hypothetical diagram chunk (currently dropped by the parser) —
diff --git a/docs/chunk_strategy/section_table.md b/docs/chunk_strategy/section_table.md
index 2b39a4e..d36fcb1 100644
--- a/docs/chunk_strategy/section_table.md
+++ b/docs/chunk_strategy/section_table.md
@@ -339,8 +339,8 @@ paragraphs>"` truncated to `section_content_max_chars`.
 {
   "chunk_type": "table",
   "table_html": "<table>...",
-  "table_text": "Cash | $4,431,280 | $7,351,900",
-  "table_text_markdown": "| Cash | $4,431,280 | $7,351,900 |\n| --- | --- | --- |",
+  "table_text": "Cash | $3,845,200 | $6,120,400",
+  "table_text_markdown": "| Cash | $3,845,200 | $6,120,400 |\n| --- | --- | --- |",
   "pre_text": "...",      // joined neighbor text (above)
   "post_text": "...",     // joined neighbor text (below)
   "token_count": <int>
@@ -444,9 +444,9 @@ Input elements (post-preprocess):
 [
   {"type": "Title",        "text": "1. ORGANIZATION",        "element_id": "h1", "metadata": {"category_depth": 0, "page_number": 7}},
   {"type": "Title",        "text": "(a) Nature of operations","element_id": "h2", "metadata": {"category_depth": 1, "page_number": 7}},
-  {"type": "NarrativeText","text": "Neksai Inc. (the \"Company\") was incorporated in Delaware on January 4, 2020.","element_id": "p1", "metadata": {"page_number": 7}},
+  {"type": "NarrativeText","text": "Contoso Inc. (the \"Company\") was incorporated in Delaware on March 12, 2019.","element_id": "p1", "metadata": {"page_number": 7}},
   {"type": "NarrativeText","text": "(Stated in U.S. Dollars (\"US$\"))","element_id": "p2", "metadata": {"page_number": 8}},
-  {"type": "Table",        "text": "Cash | $4,431,280 | $7,351,900","element_id": "t1", "metadata": {"page_number": 8, "text_as_html": "<table>...</table>"}}
+  {"type": "Table",        "text": "Cash | $3,845,200 | $6,120,400","element_id": "t1", "metadata": {"page_number": 8, "text_as_html": "<table>...</table>"}}
 ]
 ```
 
@@ -454,7 +454,7 @@ Trace:
 
 1. `Title "1. ORGANIZATION"` (depth 0) → flush (no-op), heading_path = `["1. ORGANIZATION"]`, register parent `<doc>:chunk:1`.
 2. `Title "(a) Nature of operations"` (depth 1) → flush (no-op), heading_path = `["1. ORGANIZATION", "(a) Nature of operations"]`, register parent `<doc>:chunk:2`.
-3. `NarrativeText "Neksai Inc. ..."` → append to `active_text_group`.
+3. `NarrativeText "Contoso Inc. ..."` → append to `active_text_group`.
 4. `NarrativeText "(Stated in U.S. Dollars ("US$"))"` → matches parenthetical-caption-shaped → still treated as text → append to `active_text_group`. *(Note: this only matters when the element type is `Title`/`Header`, the regex check is gated upstream. As a `NarrativeText`, it just goes to the group naturally.)*
 5. `Table` → `flush_text_group()` emits `<doc>:chunk:3` (text child of parent 2) with the two prose paragraphs joined by `\n`. Then `_build_table_chunk` collects `p2` as `pre_elements` (1 element, less than max=2; stops at Title boundary backward). No `post_elements` (end of input). Emits `<doc>:chunk:4` (table child of parent 2).
 6. End of input → final `flush_text_group()` no-op.
diff --git a/docs/chunker.md b/docs/chunker.md
index 11f46e7..d76e68d 100644
--- a/docs/chunker.md
+++ b/docs/chunker.md
@@ -241,7 +241,7 @@ Table chunks add: `table_html`, `table_text`, `table_text_markdown`, `pre_text`,
   "chunk_type": "text",
   "chunk_strategy": "title_hierarchy",
   "heading_path": ["1. ORGANIZATION", "(a) Nature of operations"],
-  "content": "1. ORGANIZATION\n(a) Nature of operations\nNeksai Inc. (the \"Company\") was incorporated ...",
+  "content": "1. ORGANIZATION\n(a) Nature of operations\nContoso Inc. (the \"Company\") was incorporated ...",
   "source_element_ids": ["t1", "t2", "n1"],
   "page_number": null,
   "parent_chunk_id": null
@@ -257,13 +257,13 @@ Table chunks add: `table_html`, `table_text`, `table_text_markdown`, `pre_text`,
   "chunk_type": "table",
   "chunk_strategy": "section_token",
   "heading_path": ["CONSOLIDATED BALANCE SHEETS"],
-  "content": "Section: CONSOLIDATED BALANCE SHEETS\n\nPre-context:\n...\n\nTable:\nCash | $4,431,280 | $7,351,900\n\nPost-context:\n...",
+  "content": "Section: CONSOLIDATED BALANCE SHEETS\n\nPre-context:\n...\n\nTable:\nCash | $3,845,200 | $6,120,400\n\nPost-context:\n...",
   "source_element_ids": ["table-element-1"],
   "page_number": 8,
   "parent_chunk_id": "avatar_comments_doc:chunk:11",
   "table_html": "<table>...</table>",
-  "table_text": "Cash | $4,431,280 | $7,351,900",
-  "table_text_markdown": "| Cash | $4,431,280 | $7,351,900 |\n| --- | --- | --- |",
+  "table_text": "Cash | $3,845,200 | $6,120,400",
+  "table_text_markdown": "| Cash | $3,845,200 | $6,120,400 |\n| --- | --- | --- |",
   "pre_text": "",
   "post_text": "Giving retroactive effect to the issuance of 1 ordinary share..."
 }
diff --git a/docs/command.md b/docs/command.md
index b692a54..282ac17 100644
--- a/docs/command.md
+++ b/docs/command.md
@@ -48,7 +48,7 @@ Example:
 uv run python -m src.cli convert \
   --source docx \
   --target pdf \
-  --input "data/Pj AVATAR_F page (260205)-with comments.docx" \
+  --input "data/ir-sample-a.docx" \
   --output data/raw/avatar-comments.pdf
 ```
 
@@ -198,7 +198,7 @@ Typical use:
 Build structured retrieval chunks from normalized parser elements.
 
 Detailed chunker behavior and output schema are documented in
-[`docs/chunker.md`](/data/lucas/projects/rag/ir_v2/docs/chunker.md).
+[`docs/chunker.md`](/docs/chunker.md).
 
 Help:
 
@@ -441,7 +441,7 @@ Behavior:
 Run one-shot baseline RAG over a local `chunks.json` file.
 
 Detailed behavior is documented in
-[`docs/rag-baseline.md`](/data/lucas/projects/rag/ir_v2/docs/rag-baseline.md).
+[`docs/rag-baseline.md`](/docs/rag-baseline.md).
 
 Help:
 
diff --git a/docs/enrichment.md b/docs/enrichment.md
index 975a12e..e8d0c58 100644
--- a/docs/enrichment.md
+++ b/docs/enrichment.md
@@ -60,12 +60,12 @@ enrichment:
 If `pipeline` is empty, indexing behaves as before and embeds plain chunk
 `content`.
 
-Ready-made configs live under [`configs/enrichment/`](/data/lucas/projects/rag/ir_v2/configs/enrichment):
+Ready-made configs live under [`configs/enrichment/`](/configs/enrichment):
 
-- [`table-context.yml`](/data/lucas/projects/rag/ir_v2/configs/enrichment/table-context.yml)
-- [`auto-keywords.yml`](/data/lucas/projects/rag/ir_v2/configs/enrichment/auto-keywords.yml)
-- [`auto-questions.yml`](/data/lucas/projects/rag/ir_v2/configs/enrichment/auto-questions.yml)
-- [`all.yml`](/data/lucas/projects/rag/ir_v2/configs/enrichment/all.yml)
+- [`table-context.yml`](/configs/enrichment/table-context.yml)
+- [`auto-keywords.yml`](/configs/enrichment/auto-keywords.yml)
+- [`auto-questions.yml`](/configs/enrichment/auto-questions.yml)
+- [`all.yml`](/configs/enrichment/all.yml)
 
 ## Preview Output
 
@@ -195,9 +195,9 @@ It does not embed text or call the vector store.
 
 ## Related Files
 
-- [`xrag/core/enrichment/`](/data/lucas/projects/rag/ir_v2/xrag/core/enrichment)
-- [`scripts/enrichment/preview.py`](/data/lucas/projects/rag/ir_v2/scripts/enrichment/preview.py)
-- [`scripts/enrichment/show_chunks.py`](/data/lucas/projects/rag/ir_v2/scripts/enrichment/show_chunks.py)
-- [`xrag/pipelines/indexing.py`](/data/lucas/projects/rag/ir_v2/xrag/pipelines/indexing.py)
-- [`docs/command.md`](/data/lucas/projects/rag/ir_v2/docs/command.md)
-- [`docs/sub-plans/ragflow-port-plan.md`](/data/lucas/projects/rag/ir_v2/docs/sub-plans/ragflow-port-plan.md)
+- [`xrag/core/enrichment/`](/xrag/core/enrichment)
+- [`scripts/enrichment/preview.py`](/scripts/enrichment/preview.py)
+- [`scripts/enrichment/show_chunks.py`](/scripts/enrichment/show_chunks.py)
+- [`xrag/pipelines/indexing.py`](/xrag/pipelines/indexing.py)
+- [`docs/command.md`](/docs/command.md)
+- [`docs/sub-plans/ragflow-port-plan.md`](/docs/sub-plans/ragflow-port-plan.md)
diff --git a/docs/indexing/table-summary-ablation.md b/docs/indexing/table-summary-ablation.md
index 122e25b..fabbf0f 100644
--- a/docs/indexing/table-summary-ablation.md
+++ b/docs/indexing/table-summary-ablation.md
@@ -120,9 +120,9 @@ Two things conspire:
    currently-noisy `table_html` (10+ pad columns, split currency cells,
    repeated multi-row headers). Inspection caught:
    - `chunk:56`: "hundreds of millions of units" vs. actual $8,810
-     (2023) / $438,533 (2024).
-   - `chunk:129`: "total lease payments of $672.2 million" vs. actual
-     ~$672K — three orders of magnitude off.
+     (2023) / $471,280 (2024).
+   - `chunk:129`: "total lease payments of $540.3 million" vs. actual
+     ~$540K — three orders of magnitude off.
    - `chunk:170`: "yen" labels on USD figures.
    - `chunk:21`: "$0.01 par value" vs. actual $0.00001 (post-subdivision).
 2. **Option B drops raw cells from the embed text for tables** — the
diff --git a/docs/library.md b/docs/library.md
index b3e8e93..d53bea4 100644
--- a/docs/library.md
+++ b/docs/library.md
@@ -38,7 +38,7 @@ client = Xrag(
     generation="openai:gpt-4.1-nano",
 )
 
-doc = await client.documents.ingest("F1-file.docx", collection="ir_docs")
+doc = await client.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 result = await client.rag.ask(
     query="What is this document about?",
     collection="ir_docs",
@@ -58,7 +58,7 @@ Runnable example:
 
 ```python
 doc = await client.documents.ingest(
-    "F1-file.docx",
+    "ir-sample-b.docx",
     collection="ir_docs",
 )
 ```
diff --git a/docs/plans.md b/docs/plans.md
index 7609143..44451c8 100644
--- a/docs/plans.md
+++ b/docs/plans.md
@@ -252,7 +252,7 @@ pipeline:
   artifact_dir: data/artifacts
 
 sources:
-  - path: "data/Pj AVATAR_F page (260205)-with comments.docx"
+  - path: "data/ir-sample-a.docx"
     doc_id: avatar_comments_doc
 
 parser:
diff --git a/docs/rag-baseline.md b/docs/rag-baseline.md
index 30ea8f7..ac89786 100644
--- a/docs/rag-baseline.md
+++ b/docs/rag-baseline.md
@@ -76,7 +76,7 @@ generation:
 
 The included config file is:
 
-- [`configs/rag-simple-baseline.yml`](/data/lucas/projects/rag/ir_v2/configs/rag-simple-baseline.yml)
+- [`configs/rag-simple-baseline.yml`](/configs/rag-simple-baseline.yml)
 
 ## Input
 
@@ -110,7 +110,7 @@ The command prints JSON:
 {
   "input_path": "data/artifacts/<run_id>/chunked/chunks.json",
   "query": "What was Cash as of December 31, 2024?",
-  "answer": "$7,351,900",
+  "answer": "$6,120,400",
   "model": "gpt-4.1-mini",
   "indexed_chunks": 129,
   "retrieved_chunks": 5,
@@ -122,7 +122,7 @@ The command prints JSON:
 
 Code lives in:
 
-- [`xrag/pipelines/rag_baseline.py`](/data/lucas/projects/rag/ir_v2/xrag/pipelines/rag_baseline.py)
+- [`xrag/pipelines/rag_baseline.py`](/xrag/pipelines/rag_baseline.py)
 
 It uses:
 
diff --git a/docs/sub-plans/chunking-plan.md b/docs/sub-plans/chunking-plan.md
index 38f15a5..709824d 100644
--- a/docs/sub-plans/chunking-plan.md
+++ b/docs/sub-plans/chunking-plan.md
@@ -95,8 +95,8 @@ Practical takeaway:
 
 Local references:
 
-- [`docs/dataset-formats.md`](/data/lucas/projects/rag/ir_v2/docs/dataset-formats.md)
-- [`docs/survey/datasets.md`](/data/lucas/projects/rag/ir_v2/docs/survey/datasets.md)
+- [`docs/dataset-formats.md`](/docs/dataset-formats.md)
+- [`docs/survey/datasets.md`](/docs/survey/datasets.md)
 
 ## Recommended Chunk Model
 
diff --git a/docs/sub-plans/eval-dataset-generation-plan.md b/docs/sub-plans/eval-dataset-generation-plan.md
index 1867806..97d17e4 100644
--- a/docs/sub-plans/eval-dataset-generation-plan.md
+++ b/docs/sub-plans/eval-dataset-generation-plan.md
@@ -580,7 +580,7 @@ Each final JSONL row should follow a stable schema like:
   "doc_id": "avatar_comments_doc",
   "chunk_strategy": "section_table",
   "question": "What was Cash as of December 31, 2024?",
-  "answer": "$7,351,900",
+  "answer": "$6,120,400",
   "answer_type": "numeric",
   "reasoning_type": "single_hop_table",
   "difficulty": "easy",
diff --git a/docs/sub-plans/llm-agent-integration.md b/docs/sub-plans/llm-agent-integration.md
index 2aa5f34..17c4042 100644
--- a/docs/sub-plans/llm-agent-integration.md
+++ b/docs/sub-plans/llm-agent-integration.md
@@ -3,7 +3,7 @@
 **Status:** draft (rewrite of 2026-04-26 Path-A plan; awaiting Phase-0 kickoff)
 **Author:** Claude (planning pass, 2026-04-27)
 **Companion docs:** [`api/apis_design.md`](api/apis_design.md), [`api/phase1.md`](api/phase1.md), [`api/phase1_review.md`](api/phase1_review.md), [`api/system.md`](api/system.md), [`ingest-module-plan.md`](ingest-module-plan.md), [`../../llm-agent/docs/langchain-compat-audit.md`](../../../llm-agent/docs/langchain-compat-audit.md)
-**Target consumer repo:** `/data/lucas/projects/rag/llm-agent` (FastAPI + Django ORM + RabbitMQ; Python 3.11; uv)
+**Target consumer repo:** `/path/to/llm-agent` (FastAPI + Django ORM + RabbitMQ; Python 3.11; uv)
 **Engine repo:** `henryle97/xrag` (this repo, ir_v2/; renamed from `henryle97/fin-rag` on 2026-04-27)
 
 ---
@@ -225,7 +225,7 @@ This test is the gate for tagging v0.1.0.
 uv build                               # → dist/xrag-0.1.0-py3-none-any.whl
 uv venv --python 3.12 /tmp/xrag-test
 /tmp/xrag-test/bin/uv pip install \
-    "/data/lucas/projects/rag/ir_v2/dist/xrag-0.1.0-py3-none-any.whl[rerank-cohere]"
+    "dist/xrag-0.1.0-py3-none-any.whl[rerank-cohere]"
 /tmp/xrag-test/bin/python -c "from xrag import run_rag_query, run_ingest_pipeline; print('ok')"
 ```
 
@@ -372,7 +372,7 @@ Self-contained, does not piggyback on `EventManagerConfig`.
 
 Reproduce the dev-run shape via the new API end-to-end:
 
-1. `POST /api/v1/rag/documents` with `data/raw/ir_document/F1-file.docx` → status reaches `ready` within 2 minutes.
+1. `POST /api/v1/rag/documents` with `data/raw/ir_document/ir-sample-b.docx` → status reaches `ready` within 2 minutes.
 2. `POST /api/v1/rag/retrievals` with the 87-question testset queries → assert top-line metrics match `experiments/eval/rag-ir-mvp3-all-docs-20260426-1658/summary.txt` within noise:
    - `deterministic.retrieval_hit ≥ 0.99`
    - `deterministic.recall_at_10 ≥ 0.97`
diff --git a/docs/sub-plans/multi-doc-rag-mvp-plan.md b/docs/sub-plans/multi-doc-rag-mvp-plan.md
index 52563c0..e7eddcc 100644
--- a/docs/sub-plans/multi-doc-rag-mvp-plan.md
+++ b/docs/sub-plans/multi-doc-rag-mvp-plan.md
@@ -335,7 +335,7 @@ Add 5–10 questions the auto-gen pipeline can't produce:
   full-eval mode.
 - **`num_match` quality.** It's a substring check, not an LLM
   judge. It will under-count cases where the model paraphrases
-  the right number ("$4.4M" vs "4,431,280"). Accept this for
+  the right number ("$3.8M" vs "3,845,200"). Accept this for
   the inner loop; full eval at promotion catches it.
 
 ### Notes
diff --git a/docs/sub-plans/parser-preprocess-plan.md b/docs/sub-plans/parser-preprocess-plan.md
index 9fd2e31..3ded850 100644
--- a/docs/sub-plans/parser-preprocess-plan.md
+++ b/docs/sub-plans/parser-preprocess-plan.md
@@ -88,7 +88,7 @@ Recommended heading heuristic:
   - repeated 3 or more times
   - short text, e.g. <= 80 chars
   - not a substantive sentence
-  - common company/page-header style content such as `NEKSAI INC.`
+  - common company/page-header style content such as `CONTOSO INC.`
 
 6. `prefer_table_html(elements)`
 
diff --git a/docs/sub-plans/retrieval-quality-fixes.md b/docs/sub-plans/retrieval-quality-fixes.md
index 5fe0136..33f7892 100644
--- a/docs/sub-plans/retrieval-quality-fixes.md
+++ b/docs/sub-plans/retrieval-quality-fixes.md
@@ -14,8 +14,8 @@ defaults: `gpt-4.1-nano` generator, 3 RAGAS metrics, `gpt-4o-mini` judge).
 
 | Trace | Query | Expected | Answer | Primary source rank |
 |---|---|---|---|---|
-| 1 | "total rent expenditure for FY 2023" | US$298,754 | ✅ correct | **1** (lucky) |
-| 2 | "total cash by December 31, 2024" | US$7,351,900 | ✅ correct | **8** (balance sheet chunk 12) |
+| 1 | "total rent expenditure for FY 2023" | US$341,620 | ✅ correct | **1** (lucky) |
+| 2 | "total cash by December 31, 2024" | US$6,120,400 | ✅ correct | **8** (balance sheet chunk 12) |
 
 Answers are correct, but the ranking reveals three problems that will
 bite harder as we drop to smaller top-k, introduce reranking, or move
@@ -67,7 +67,7 @@ Two compounding causes:
    has ~40 rows covering 4 line-item categories. BM25's length
    normalisation penalises long docs; the dense embedding captures a
    diffuse "financial statement" topic vector rather than a specific
-   "cash" signal. The answer row ("Cash | $4,431,280 | $7,351,900") is
+   "cash" signal. The answer row ("Cash | $3,845,200 | $6,120,400") is
    one of dozens.
 
 Today the generator gets the right answer because a *different* chunk
diff --git a/docs/sub-plans/section-table-chunker-review.md b/docs/sub-plans/section-table-chunker-review.md
index 66193f8..9a92f50 100644
--- a/docs/sub-plans/section-table-chunker-review.md
+++ b/docs/sub-plans/section-table-chunker-review.md
@@ -312,8 +312,8 @@ Mapping items in
 | Bucket | Item | Chunker contribution |
 |---|---|---|
 | C — wrong-cell (×14) | `ETH 2023 = 290` | S3: gold in chunk:120 (cont.), retrieval saturated by chunks 116-118 under base parent |
-| C — wrong-cell | `salary = $(361,993)` | S3 + S2: same fragmentation; nearby paragraph emitted 3× crowds top-k |
-| D — retrieval miss (×3) | `digital assets 2023 = $5,868,425` (gold chunk:12 in balance sheet) | S3: section 114+120 carry strong "digital assets" lexical match and out-rank the balance-sheet line |
+| C — wrong-cell | `salary = $(405,760)` | S3 + S2: same fragmentation; nearby paragraph emitted 3× crowds top-k |
+| D — retrieval miss (×3) | `digital assets 2023 = $6,214,730` (gold chunk:12 in balance sheet) | S3: section 114+120 carry strong "digital assets" lexical match and out-rank the balance-sheet line |
 | E — found at low rank (×16) | `cash 2024` etc. | S1 + S6: parent/sibling near-duplicates eat top-3 slots, evidence pushed to rank 3-7 |
 
 S3 (`(cont.)` fragmentation) is the single highest-leverage finding —
diff --git a/docs/sub-plans/xrag-public-api.md b/docs/sub-plans/xrag-public-api.md
index 67bd65e..828a2e8 100644
--- a/docs/sub-plans/xrag-public-api.md
+++ b/docs/sub-plans/xrag-public-api.md
@@ -106,7 +106,7 @@ Concrete patterns observed in current docs (April 2026), via Context7. Survey is
 from xrag import Xrag
 
 client = Xrag.from_env()
-await client.documents.ingest("F1-file.docx", collection="ir_docs")
+await client.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 result = await client.rag.ask("What's the FY-23 revenue?", collection="ir_docs")
 print(result.answer)
 ```
@@ -173,14 +173,14 @@ Only `qdrant_url` and `embedding` are strictly required. Everything else has sen
 
 ```python
 # Single-tenant prototype — implicit "default" tenant
-await client.documents.ingest("F1-file.docx", collection="ir_docs")
+await client.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 
 # Explicit per-call (multi-tenant from a request handler)
-await client.documents.ingest("F1-file.docx", collection="ir_docs", tenant_id="dev")
+await client.documents.ingest("ir-sample-b.docx", collection="ir_docs", tenant_id="dev")
 
 # Bound tenant for a sequence of calls (request-scoped)
 tenant = client.for_tenant("dev")    # XragTenant — tenant_id pre-bound
-await tenant.documents.ingest("F1-file.docx", collection="ir_docs")
+await tenant.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 await tenant.retrievals.search(query="...", collection="ir_docs", top_k=10)
 ```
 
@@ -484,7 +484,7 @@ User-supplied `artifacts_dir` paths can be reused across calls — that's the wh
     "hostname": "ingest-worker-3",
     "started_at": "2026-04-27T08:14:22.481Z",
     "document_id": "doc_01HXY...",
-    "source_name": "F1-file.docx",
+    "source_name": "ir-sample-b.docx",
     "xrag_version": "0.2.0"
   }
   ```
@@ -665,7 +665,7 @@ client = Xrag(
 ```python
 # ingest as usual
 await client.documents.ingest(
-    "F1-file.docx", collection="ir_docs", tenant_id="dev",
+    "ir-sample-b.docx", collection="ir_docs", tenant_id="dev",
 )
 
 # retrieve, then generate with whatever you want
@@ -699,7 +699,7 @@ What the caller controls themselves: prompt, model choice, temperature, structur
 from xrag import XragSync
 
 client = XragSync.from_env()
-doc = client.documents.ingest("F1-file.docx", collection="ir_docs", tenant_id="dev")
+doc = client.documents.ingest("ir-sample-b.docx", collection="ir_docs", tenant_id="dev")
 print(client.rag.ask(query="...", collection="ir_docs", tenant_id="dev").answer)
 ```
 
@@ -936,7 +936,7 @@ query_result = run_rag("What's the revenue?", cfg)
 from xrag import Xrag
 
 client = Xrag.from_env()
-await client.documents.ingest("F1-file.docx", collection="ir_docs")
+await client.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 result = await client.rag.ask(query="What's the revenue?", collection="ir_docs")
 ```
 
@@ -945,7 +945,7 @@ result = await client.rag.ask(query="What's the revenue?", collection="ir_docs")
 from xrag import Xrag
 
 client = Xrag(qdrant_url="http://qdrant:6333", embedding="openai:text-embedding-3-small")
-await client.documents.ingest("F1-file.docx", collection="ir_docs")
+await client.documents.ingest("ir-sample-b.docx", collection="ir_docs")
 search = await client.retrievals.search(
     query="What's the revenue?", collection="ir_docs", top_k=10,
 )
diff --git a/docs/survey/SOTA-RAG-for-financial-documents.md b/docs/survey/SOTA-RAG-for-financial-documents.md
index e8b406f..bb3da8c 100644
--- a/docs/survey/SOTA-RAG-for-financial-documents.md
+++ b/docs/survey/SOTA-RAG-for-financial-documents.md
@@ -424,4 +424,4 @@ When you're ready for multi-company (Level 2), you download Microsoft and Tesla
 
 **You never index FinanceBench itself.** It's purely evaluation data — like how you never put your unit test assertions into your application database.
 
-Does that clear it up?
\ No newline at end of file
+Does that clear it up?
diff --git a/docs/survey/claude-rag-survey.md b/docs/survey/claude-rag-survey.md
index d3680a8..9a5bc98 100644
--- a/docs/survey/claude-rag-survey.md
+++ b/docs/survey/claude-rag-survey.md
@@ -11,7 +11,7 @@ Building a RAG system for financial reports (DOCX/PDF) with dense text + many ta
 
 ## CHALLENGE 1: Document Parsing & Table Extraction
 
-**Problem**: Financial reports like the Neksai Inc. filing contain 20+ complex tables (balance sheets, income statements, cash flows, footnote tables) interleaved with dense narrative text. Standard text extraction destroys table structure, making retrieval of exact numbers impossible.
+**Problem**: Financial reports like the Contoso Inc. filing contain 20+ complex tables (balance sheets, income statements, cash flows, footnote tables) interleaved with dense narrative text. Standard text extraction destroys table structure, making retrieval of exact numbers impossible.
 
 ### Solution A: MinerU + HTML Table Preservation
 
@@ -51,7 +51,7 @@ Step 1.6  Validate extraction against golden samples:
 
 ## CHALLENGE 2: Table-Aware Chunking Strategy
 
-**Problem**: Standard fixed-size chunking (400–512 tokens) splits tables mid-row, destroying the relationship between row headers and values. A question like "What was Neksai's total current assets in 2024?" requires the entire balance sheet table to be in one chunk, but that table may be 800+ tokens.
+**Problem**: Standard fixed-size chunking (400–512 tokens) splits tables mid-row, destroying the relationship between row headers and values. A question like "What was Contoso's total current assets in 2024?" requires the entire balance sheet table to be in one chunk, but that table may be 800+ tokens.
 
 ### Solution A: Structure-Aware Chunking with Table Atomization
 
@@ -76,9 +76,9 @@ Chunk types:
 
 ```
 Parent: Full "Balance Sheet" section (table + surrounding text)
-  └─ Child: "Total current assets: $13,387,879 (Dec 31, 2024)"
-  └─ Child: "Cash: $7,351,900 (Dec 31, 2024)"
-  └─ Child: "Digital assets: $2,853,590 (Dec 31, 2024)"
+  └─ Child: "Total current assets: $15,640,290 (Dec 31, 2024)"
+  └─ Child: "Cash: $6,120,400 (Dec 31, 2024)"
+  └─ Child: "Digital assets: $3,106,470 (Dec 31, 2024)"
 ```
 
 - **Pros**: Precise retrieval + complete context, solves the fundamental chunk-size tradeoff (RAGFlow's "structural conflict"), enables multi-hop reasoning
@@ -99,11 +99,11 @@ Step 2.1  Define chunk hierarchy:
 Step 2.2  For each table, generate BOTH:
           a) Full table as HTML (parent chunk)
           b) Row-group summaries as natural language (child chunks):
-             "As of December 31, 2024, Neksai's cash was $7,351,900 and
-              digital assets were $2,853,590"
+             "As of December 31, 2024, Contoso's cash was $6,120,400 and
+              digital assets were $3,106,470"
 Step 2.3  Add contextual metadata to every chunk:
           {
-            "doc_id": "neksai_10k_2024",
+            "doc_id": "contoso_10k_2024",
             "section": "Consolidated Balance Sheets",
             "page": 3,
             "chunk_type": "table",
@@ -120,7 +120,7 @@ Step 2.5  Validate: for each gold Q&A pair, verify the answer exists in at
 
 ## CHALLENGE 3: Hybrid Retrieval + Reranking
 
-**Problem**: Dense vector search alone misses exact financial terms (e.g., "PCAOB ID:1171", "ASU 2023-08", specific dollar amounts). Sparse search alone misses semantic paraphrases ("profit" vs. "net income"). Financial queries require both.
+**Problem**: Dense vector search alone misses exact financial terms (e.g., "PCAOB ID:9999", "ASU 2023-08", specific dollar amounts). Sparse search alone misses semantic paraphrases ("profit" vs. "net income"). Financial queries require both.
 
 ### Solution A: BM25 + Dense Vectors + Reciprocal Rank Fusion (RRF)
 
@@ -234,10 +234,10 @@ Step 4.1  Design generation prompt with:
           - Uncertainty handling ("INSUFFICIENT_CONTEXT")
           - Output format: {answer, citations[], confidence, reasoning}
 Step 4.2  Implement answer type detection:
-          - EXACT_VALUE: "What was total revenue in 2024?" → $45,042,408
+          - EXACT_VALUE: "What was total revenue in 2024?" → $48,715,220
           - COMPARISON: "How did revenue change YoY?" → calculation + values
-          - EXPLANATION: "What is Neksai's revenue recognition policy?" → text
-          - YES_NO: "Does Neksai have deferred tax assets?" → Yes/No + evidence
+          - EXPLANATION: "What is Contoso's revenue recognition policy?" → text
+          - YES_NO: "Does Contoso have deferred tax assets?" → Yes/No + evidence
 Step 4.3  For EXACT_VALUE questions:
           - Extract the number from the answer
           - Regex-validate it exists in a retrieved chunk
@@ -351,19 +351,19 @@ Evaluation layers:
 ### Implementation Steps
 
 ```
-Step 6.1  Build gold QA dataset from the Neksai financial statements:
+Step 6.1  Build gold QA dataset from the Contoso financial statements:
           Category 1 — Exact Value (deterministic):
-            Q: "What was Neksai's net income for 2024?"
-            A: "$5,741,797"
+            Q: "What was Contoso's net income for 2024?"
+            A: "$6,003,510"
             Gold chunks: [balance_sheet_2024, income_statement_2024]
 
           Category 2 — Calculation:
             Q: "What was the gross profit margin in 2024?"
-            A: "64.1% ($28,851,979 / $45,042,408)"
+            A: "64.1% ($31,402,650 / $48,715,220)"
             Gold chunks: [income_statement_2024]
 
           Category 3 — Explanation (LLM-judged):
-            Q: "How does Neksai recognize revenue from subscription services?"
+            Q: "How does Contoso recognize revenue from subscription services?"
             A: "Over time within the subscription period..."
             Gold chunks: [note_2_revenue_recognition]
 
@@ -374,7 +374,7 @@ Step 6.1  Build gold QA dataset from the Neksai financial statements:
             Gold chunks: [note_1_reorganization]
 
           Category 5 — Unanswerable:
-            Q: "What is Neksai's projected revenue for 2025?"
+            Q: "What is Contoso's projected revenue for 2025?"
             A: "INSUFFICIENT_CONTEXT"
             Gold chunks: [] (nothing should be relevant)
 
@@ -488,7 +488,7 @@ Step 6.5  Private-set robustness measures:
 ### Phase 1: Foundation (Weeks 1-2)
 | Step | Task | Deliverable |
 |------|------|-------------|
-| 1.1 | Set up MinerU + Docling parsing | Parsed Neksai report with 95%+ table accuracy |
+| 1.1 | Set up MinerU + Docling parsing | Parsed Contoso report with 95%+ table accuracy |
 | 1.2 | Build multi-granularity chunker | Parent/child chunk pairs with metadata |
 | 1.3 | Set up vector store + BM25 index | Indexed chunks, searchable |
 | 1.4 | Basic retrieval pipeline (dense only) | Baseline Recall@5 measurement |
@@ -559,4 +559,4 @@ Step 6.5  Private-set robustness measures:
 - **GaRAGe** (ACL 2025) — Grounding annotations benchmark for RAG evaluation
 - **RAGAS** — Automated RAG evaluation framework
 - **MinerU** (OpenDataLab) — 97.5 mAP layout detection, best table extraction
-- **Docling** (IBM) — MIT-licensed, TableFormer-powered parsing
\ No newline at end of file
+- **Docling** (IBM) — MIT-licensed, TableFormer-powered parsing
diff --git a/docs/survey/datasets.md b/docs/survey/datasets.md
index ca3ced8..fc7046c 100644
--- a/docs/survey/datasets.md
+++ b/docs/survey/datasets.md
@@ -1,6 +1,6 @@
 # Datasets
 
-- **FinanceBench** 
+- **FinanceBench**
 - **T2-RAGBench**
 - FinDER
 Here's the landscape of RAG evaluation datasets, organized from most to least relevant for your financial report use case:
@@ -41,6 +41,6 @@ Here's the landscape of RAG evaluation datasets, organized from most to least re
 
 **FiQA** (2018) — opinion-based financial QA, mainly sentiment. **ConvFinQA** (2022) — conversational multi-turn financial QA. **HotpotQA** — general multi-hop, not financial. **NaturalQuestions** — general single-hop from Google search.
 
-## What to use for your Neksai project
+## What to use for your Contoso project
 
 Start with **FinanceBench** as your primary reference for format and question style — its evidence provenance schema (evidence_text + page + document) directly inspired your golden dataset design. Use **T2-RAGBench** or **FinQA** as a secondary reference for table-heavy numerical reasoning evaluation. For multi-hop testing, look at **FinDER** for realistic analyst-style queries. Your custom dataset fills the gap that none of these cover: single-company, full-report RAG with deterministic + LLM-judge hybrid evaluation and mandatory telemetry scoring.
diff --git a/docs/survey/grok-rag-survey.md b/docs/survey/grok-rag-survey.md
index e2471b8..6e735b0 100644
--- a/docs/survey/grok-rag-survey.md
+++ b/docs/survey/grok-rag-survey.md
@@ -61,6 +61,6 @@ For a deployable system today (English SEC/earnings reports):
 - Add post-retrieval re-ranking and citation/attribution (cell-level in HierFinRAG).
 - For scale: Binary quantization + hybrid search reduces costs dramatically.
 
-These approaches (especially HierFinRAG + Unstructured.io backbone) represent the current English-language SOTA for reliable, hallucination-resistant RAG on text + table financial reports. They move far beyond vanilla RAG by explicitly modeling the interplay between narrative and numerical data. 
+These approaches (especially HierFinRAG + Unstructured.io backbone) represent the current English-language SOTA for reliable, hallucination-resistant RAG on text + table financial reports. They move far beyond vanilla RAG by explicitly modeling the interplay between narrative and numerical data.
 
-If you need code snippets, a specific implementation guide (e.g., LlamaIndex + Unstructured + HierFinRAG components), or evaluation on a sample 10-K, let me know!
\ No newline at end of file
+If you need code snippets, a specific implementation guide (e.g., LlamaIndex + Unstructured + HierFinRAG components), or evaluation on a sample 10-K, let me know!
diff --git a/docs/survey/preprocessing-and-normalization.md b/docs/survey/preprocessing-and-normalization.md
index 283acb8..2872e42 100644
--- a/docs/survey/preprocessing-and-normalization.md
+++ b/docs/survey/preprocessing-and-normalization.md
@@ -130,8 +130,8 @@ The largest ROI lever for financial documents.
   headers ("Year ended December 31" spanning 2024/2023).
 - **Empty row / column trimming** — visual spacers between sections
   produce all-empty rows; drop them.
-- **Currency-cell merging** — accounting style `$ | 4,431,280` should
-  become a single cell `$4,431,280`; otherwise the `$` becomes its own
+- **Currency-cell merging** — accounting style `$ | 3,845,200` should
+  become a single cell `$3,845,200`; otherwise the `$` becomes its own
   noise column. (This repo's `merge_table_currency_cells` step.)
 - **Section-row preservation** — `ASSETS`, `Current assets:` headers
   should remain as anchor rows, not be dropped as "empty after the
@@ -328,7 +328,7 @@ Specific failure modes preprocessing should improve:
   because heading was a `UncategorizedText`. Caught by recall drop on
   questions whose answers live in body content but are
   heading-disambiguated.
-- **Table currency mis-segmentation** — `$ | 4,431,280` → two cells →
+- **Table currency mis-segmentation** — `$ | 3,845,200` → two cells →
   retriever returns either the `$` cell or the number cell; generator
   hallucinates the missing half. Caught by num_match drops on
   numeric-answer questions.
diff --git a/examples/01_quickstart.py b/examples/01_quickstart.py
index 3551b54..c2c03ad 100644
--- a/examples/01_quickstart.py
+++ b/examples/01_quickstart.py
@@ -15,7 +15,7 @@
 
 from xrag import Xrag
 
-DOC_PATH = Path("data/raw/ir_document/F1-file.docx")
+DOC_PATH = Path("data/raw/ir_document/ir-sample-b.docx")
 COLLECTION = "ir_quickstart"
 
 
diff --git a/examples/04_multi_tenant.py b/examples/04_multi_tenant.py
index 1154af7..235c6aa 100644
--- a/examples/04_multi_tenant.py
+++ b/examples/04_multi_tenant.py
@@ -11,8 +11,8 @@
 
 from xrag import Xrag
 
-DOC_A = Path("data/raw/ir_document/F1-file.docx")
-DOC_B = Path("data/raw/ir_document/Pj_AVATAR_F_page_260205.docx")
+DOC_A = Path("data/raw/ir_document/ir-sample-b.docx")
+DOC_B = Path("data/raw/ir_document/ir-sample-a.docx")
 COLLECTION = "ir_multitenant"
 
 
diff --git a/examples/08_per_call_overrides.py b/examples/08_per_call_overrides.py
index b421323..4bb66b7 100644
--- a/examples/08_per_call_overrides.py
+++ b/examples/08_per_call_overrides.py
@@ -12,7 +12,7 @@
 
 from xrag import Xrag
 
-DOC_PATH = Path("data/raw/ir_document/F1-file.docx")
+DOC_PATH = Path("data/raw/ir_document/ir-sample-b.docx")
 COLLECTION = "ir_overrides"
 
 
diff --git a/examples/09_parse_docx_api_vs_local.py b/examples/09_parse_docx_api_vs_local.py
index 9b8a9c4..5394e50 100644
--- a/examples/09_parse_docx_api_vs_local.py
+++ b/examples/09_parse_docx_api_vs_local.py
@@ -25,7 +25,7 @@
 from xrag.config.models import AppConfig, PipelineConfig, ProviderConfig, SourceConfig
 from xrag.pipelines.parser import run_parser_pipeline
 
-DOC_PATH = Path("data/raw/ir_document/F1-file.docx")
+DOC_PATH = Path("data/raw/ir_document/ir-sample-b.docx")
 ROOT = Path("data/dev/xrag_examples/09_parse_docx_api_vs_local")
 
 
diff --git a/pyproject.toml b/pyproject.toml
index e616dc9..d792d28 100644
--- a/pyproject.toml
+++ b/pyproject.toml
@@ -93,8 +93,6 @@ select = ["E", "F", "I", "UP", "B"]
 "tools/review/chunker_review.py" = ["E501"]
 "tools/eval_dataset/review.py" = ["E501"]
 "tools/evaluation/preview.py" = ["E501"]
-# news.py is a standalone script, not part of the package
-"news.py" = ["E501", "B904"]
 
 [tool.ruff.format]
 quote-style = "double"
diff --git a/scripts/ingest/ir_latest.sh b/scripts/ingest/ir_latest.sh
index 0a8f158..60abc4d 100755
--- a/scripts/ingest/ir_latest.sh
+++ b/scripts/ingest/ir_latest.sh
@@ -17,10 +17,10 @@ REPO_ROOT="$(cd "$(dirname "${BASH_SOURCE[0]}")/../.." && pwd)"
 cd "$REPO_ROOT"
 
 CONFIG="configs/rag-ir-latest.yml"
-RAW_DIR="/data/lucas/projects/rag/ir_v2/data/dev/ir_document/raw"
-OUTPUT_DIR="/data/lucas/projects/rag/ir_v2/data/dev/ir_document/latest"
-F1_DOC="$RAW_DIR/F1-file.docx"
-PJ_DOC="$RAW_DIR/Pj_AVATAR_F_page_260205.docx"
+RAW_DIR="data/dev/ir_document/raw"
+OUTPUT_DIR="data/dev/ir_document/latest"
+F1_DOC="$RAW_DIR/ir-sample-b.docx"
+PJ_DOC="$RAW_DIR/ir-sample-a.docx"
 
 if [[ ! -f "$CONFIG" ]]; then
   echo "ERROR: config not found: $CONFIG" >&2
diff --git a/scripts/parser/parse_f1_doc.sh b/scripts/parser/parse_f1_doc.sh
index 565db4c..2ec01f9 100644
--- a/scripts/parser/parse_f1_doc.sh
+++ b/scripts/parser/parse_f1_doc.sh
@@ -128,4 +128,3 @@ PY
 
 # uv run python -m xrag.cli parser \
 #     --config data/dev/ir_document/configs/parsers/parse-f1-doc.yml
-
diff --git a/scripts/parser/parse_pj_doc.sh b/scripts/parser/parse_pj_doc.sh
index d124ffd..f6e50e9 100644
--- a/scripts/parser/parse_pj_doc.sh
+++ b/scripts/parser/parse_pj_doc.sh
@@ -126,4 +126,3 @@ PY
 
 # uv run python -m xrag.cli parser \
 #     --config data/dev/ir_document/configs/parsers/parse-f1-doc.yml
-
diff --git a/tests/unit/test_chunk_prepare.py b/tests/unit/test_chunk_prepare.py
index 87b4722..6e88b02 100644
--- a/tests/unit/test_chunk_prepare.py
+++ b/tests/unit/test_chunk_prepare.py
@@ -36,13 +36,13 @@ def test_run_chunk_prepare_pipeline_builds_section_text_and_table_chunks(tmp_pat
                     },
                     {
                         "type": "Table",
-                        "text": "Cash | $4,431,280 | $7,351,900",
+                        "text": "Cash | $3,845,200 | $6,120,400",
                         "element_id": "table-1",
                         "metadata": {
                             "page_number": 3,
                             "text_as_html": (
-                                "<table><tr><td>Cash</td><td>$4,431,280</td>"
-                                "<td>$7,351,900</td></tr></table>"
+                                "<table><tr><td>Cash</td><td>$3,845,200</td>"
+                                "<td>$6,120,400</td></tr></table>"
                             ),
                         },
                     },
@@ -76,9 +76,9 @@ def test_run_chunk_prepare_pipeline_builds_section_text_and_table_chunks(tmp_pat
     assert text_chunk["parent_id"] == section_chunk["id"]
     assert table_chunk["parent_id"] == section_chunk["id"]
     assert table_chunk["metadata"]["table_html"].startswith("<table>")
-    assert table_chunk["metadata"]["table_text"] == "Cash | $4,431,280 | $7,351,900"
+    assert table_chunk["metadata"]["table_text"] == "Cash | $3,845,200 | $6,120,400"
     assert table_chunk["metadata"]["table_text_markdown"] == (
-        "| Cash | $4,431,280 | $7,351,900 |\n| --- | --- | --- |"
+        "| Cash | $3,845,200 | $6,120,400 |\n| --- | --- | --- |"
     )
     assert (
         "The balance sheets present the financial position." in table_chunk["metadata"]["pre_text"]
@@ -153,13 +153,13 @@ def test_run_chunk_prepare_pipeline_supports_section_token_strategy(tmp_path: Pa
                     },
                     {
                         "type": "Table",
-                        "text": "Cash | $4,431,280 | $7,351,900",
+                        "text": "Cash | $3,845,200 | $6,120,400",
                         "element_id": "table-1",
                         "metadata": {
                             "page_number": 3,
                             "text_as_html": (
-                                "<table><tr><td>Cash</td><td>$4,431,280</td>"
-                                "<td>$7,351,900</td></tr></table>"
+                                "<table><tr><td>Cash</td><td>$3,845,200</td>"
+                                "<td>$6,120,400</td></tr></table>"
                             ),
                         },
                     },
@@ -196,7 +196,7 @@ def test_run_chunk_prepare_pipeline_supports_section_token_strategy(tmp_path: Pa
     assert table_chunk["metadata"]["chunk_strategy"] == "section_token"
     assert table_chunk["parent_id"] == section_chunk["id"]
     assert table_chunk["metadata"]["table_text_markdown"].startswith(
-        "| Cash | $4,431,280 | $7,351,900 |"
+        "| Cash | $3,845,200 | $6,120,400 |"
     )
     assert table_chunk["metadata"]["pre_text"]
     assert table_chunk["metadata"]["post_text"]
@@ -323,13 +323,13 @@ def test_section_table_chunker_skips_parenthetical_caption_as_heading(
                 },
                 {
                     "type": "Table",
-                    "text": "Cash | 4,431,280 | 7,351,900",
+                    "text": "Cash | 3,845,200 | 6,120,400",
                     "element_id": "table-1",
                     "metadata": {
                         "page_number": 3,
                         "text_as_html": (
-                            "<table><tr><td>Cash</td><td>4,431,280</td>"
-                            "<td>7,351,900</td></tr></table>"
+                            "<table><tr><td>Cash</td><td>3,845,200</td>"
+                            "<td>6,120,400</td></tr></table>"
                         ),
                     },
                 },
diff --git a/tests/unit/test_embed_text.py b/tests/unit/test_embed_text.py
index 3ad4e8d..c73593c 100644
--- a/tests/unit/test_embed_text.py
+++ b/tests/unit/test_embed_text.py
@@ -89,7 +89,7 @@ def test_table_chunk_with_summary_and_questions_orders_table_text_first() -> Non
     """
     chunk = {
         "chunk_type": "table",
-        "table_text": "Cash | $4,431,280 | $7,351,900",
+        "table_text": "Cash | $3,845,200 | $6,120,400",
         "summary_kwd": "Balance sheet snapshot of cash position year over year.",
         "question_kwd": [
             "What was Cash on the December 2025 balance sheet?",
@@ -99,7 +99,7 @@ def test_table_chunk_with_summary_and_questions_orders_table_text_first() -> Non
     }
 
     assert compose_embed_text(chunk) == (
-        "Cash | $4,431,280 | $7,351,900\n\n"
+        "Cash | $3,845,200 | $6,120,400\n\n"
         "Balance sheet snapshot of cash position year over year.\n\n"
         "Questions:\n"
         "What was Cash on the December 2025 balance sheet?\n"
diff --git a/tests/unit/test_enrichment.py b/tests/unit/test_enrichment.py
index 79c072c..69cb8da 100644
--- a/tests/unit/test_enrichment.py
+++ b/tests/unit/test_enrichment.py
@@ -53,8 +53,8 @@ def test_run_enrichment_pipeline_applies_ragflow_fields(monkeypatch) -> None:
             "chunk_id": "c2",
             "chunk_type": "table",
             "heading_path": ["CONSOLIDATED BALANCE SHEETS"],
-            "content": "Cash 4,431,280 7,351,900",
-            "table_text": "Cash | 4,431,280 | 7,351,900",
+            "content": "Cash 3,845,200 6,120,400",
+            "table_text": "Cash | 3,845,200 | 6,120,400",
         },
         {
             "chunk_id": "c3",
diff --git a/tests/unit/test_eval_create_dataset.py b/tests/unit/test_eval_create_dataset.py
index 720652e..f281896 100644
--- a/tests/unit/test_eval_create_dataset.py
+++ b/tests/unit/test_eval_create_dataset.py
@@ -18,8 +18,8 @@
 _SIMPLE_TABLE_HTML = (
     "<table>"
     "<tr><td></td><td>December 31, 2023</td><td>December 31, 2024</td></tr>"
-    "<tr><td>Cash</td><td>$4,431,280</td><td>$7,351,900</td></tr>"
-    "<tr><td>Digital assets</td><td>$317,662</td><td>$1,012,543</td></tr>"
+    "<tr><td>Cash</td><td>$3,845,200</td><td>$6,120,400</td></tr>"
+    "<tr><td>Digital assets</td><td>$352,940</td><td>$1,012,543</td></tr>"
     "</table>"
 )
 
@@ -42,7 +42,7 @@ def _build_table_chunk(
         "page_number": page_number,
         "parent_chunk_id": "test_doc:chunk:1",
         "table_html": table_html,
-        "table_text": "Cash | $4,431,280 | $7,351,900",
+        "table_text": "Cash | $3,845,200 | $6,120,400",
         "pre_text": "The balance sheets present the financial position.",
         "post_text": "Cash increased year over year.",
     }
@@ -167,7 +167,7 @@ def test_parse_table_html_extracts_grid() -> None:
     assert "Cash" in grid.row_labels
     assert "Digital assets" in grid.row_labels
     assert len(grid.values) == 2
-    assert any("4,431,280" in v for v in grid.values[0])
+    assert any("3,845,200" in v for v in grid.values[0])
 
 
 def test_parse_table_html_returns_none_for_empty() -> None:
diff --git a/tests/unit/test_eval_dataset_review.py b/tests/unit/test_eval_dataset_review.py
index a22d4a8..1bad6f4 100644
--- a/tests/unit/test_eval_dataset_review.py
+++ b/tests/unit/test_eval_dataset_review.py
@@ -19,7 +19,7 @@ def _sample_items() -> list[dict]:
         {
             "example_id": "doc:qa:0001",
             "question": "What was total cash as of Dec 31?",
-            "answer": "$7,351,900",
+            "answer": "$6,120,400",
             "answer_type": "numeric",
             "reasoning_type": "table_lookup",
             "difficulty": "easy",
@@ -123,7 +123,7 @@ def test_load_chunks_index_handles_canonical_envelope(tmp_path: Path) -> None:
                     {
                         "id": "doc:chunk:12",
                         "doc_id": "doc",
-                        "text": "Cash at Dec 31 was 7351900.",
+                        "text": "Cash at Dec 31 was 6120400.",
                         "metadata": {
                             "heading_path": ["Balance Sheet"],
                         },
@@ -138,7 +138,7 @@ def test_load_chunks_index_handles_canonical_envelope(tmp_path: Path) -> None:
 
     index = load_chunks_index(chunks_path)
 
-    assert index["doc:chunk:12"]["content"] == "Cash at Dec 31 was 7351900."
+    assert index["doc:chunk:12"]["content"] == "Cash at Dec 31 was 6120400."
     assert index["doc:chunk:12"]["heading_path"] == ["Balance Sheet"]
 
 
@@ -165,7 +165,7 @@ def test_generate_dataset_review_html_with_chunks_resolves_evidence() -> None:
     chunks_by_id = {
         "doc:chunk:12": {
             "chunk_id": "doc:chunk:12",
-            "content": "Cash at Dec 31 was 7351900.",
+            "content": "Cash at Dec 31 was 6120400.",
             "heading_path": ["Balance Sheet"],
             "chunk_type": "table",
         }
@@ -173,7 +173,7 @@ def test_generate_dataset_review_html_with_chunks_resolves_evidence() -> None:
     html_out = generate_dataset_review_html(items, chunks_by_id=chunks_by_id)
 
     # The chunk content must be embedded somewhere so the JS can resolve it
-    assert "Cash at Dec 31 was 7351900." in html_out
+    assert "Cash at Dec 31 was 6120400." in html_out
     assert "Balance Sheet" in html_out
 
 
@@ -207,7 +207,7 @@ def test_eval_review_cli_with_chunks(tmp_path: Path) -> None:
                 "chunks": [
                     {
                         "chunk_id": "doc:chunk:12",
-                        "content": "Cash at Dec 31 was 7351900.",
+                        "content": "Cash at Dec 31 was 6120400.",
                         "heading_path": ["Balance Sheet"],
                         "chunk_type": "table",
                     }
@@ -234,4 +234,4 @@ def test_eval_review_cli_with_chunks(tmp_path: Path) -> None:
 
     assert result.exit_code == 0, result.stdout
     rendered = output_path.read_text()
-    assert "Cash at Dec 31 was 7351900." in rendered
+    assert "Cash at Dec 31 was 6120400." in rendered
diff --git a/tests/unit/test_eval_run.py b/tests/unit/test_eval_run.py
index 7fee505..93aab0a 100644
--- a/tests/unit/test_eval_run.py
+++ b/tests/unit/test_eval_run.py
@@ -126,7 +126,7 @@ def _build_testset_jsonl(tmp_path: Path) -> Path:
             "example_id": "doc1:qa:0001",
             "doc_id": "doc1",
             "question": "What was Cash as of 2024?",
-            "answer": "$7,351,900",
+            "answer": "$6,120,400",
             "answer_type": "numeric",
             "reasoning_type": "single_hop_table",
             "difficulty": "easy",
@@ -226,10 +226,10 @@ def test_deterministic_metrics_numeric_match() -> None:
     q = EvalQuestion(
         question_id="q1",
         question="Q?",
-        ground_truth="$4,431,280",
+        ground_truth="$3,845,200",
         answer_type="numeric",
     )
-    metrics = _compute_deterministic_metrics(q, "The value was 4431280 dollars.", [])
+    metrics = _compute_deterministic_metrics(q, "The value was 3845200 dollars.", [])
     assert metrics["numeric_match"] == 1.0
 
 
@@ -248,7 +248,7 @@ def test_deterministic_metrics_no_evidence_skips_retrieval() -> None:
 
 def test_extract_number_handles_formatting() -> None:
     """Extract number from various text formats."""
-    assert _extract_number("$4,431,280") == 4431280.0
+    assert _extract_number("$3,845,200") == 3845200.0
     assert _extract_number("about 3.5 million") == 3.5
     assert _extract_number("no number here") is None
 
diff --git a/tests/unit/test_llm_dataset_generator.py b/tests/unit/test_llm_dataset_generator.py
index f78769a..0fa80a6 100644
--- a/tests/unit/test_llm_dataset_generator.py
+++ b/tests/unit/test_llm_dataset_generator.py
@@ -67,13 +67,13 @@ def _make_table_chunk(chunk_id: str = "test_doc:chunk:10") -> dict[str, Any]:
     return _make_chunk(
         chunk_id=chunk_id,
         chunk_type="table",
-        content="Cash | $4,431,280 | $7,351,900",
+        content="Cash | $3,845,200 | $6,120,400",
         heading_path=["CONSOLIDATED BALANCE SHEETS"],
         table_html=(
             "<table>"
             "<tr><td></td><td>December 31, 2023</td><td>December 31, 2024</td></tr>"
-            "<tr><td>Cash</td><td>$4,431,280</td><td>$7,351,900</td></tr>"
-            "<tr><td>Digital assets</td><td>$317,662</td><td>$1,012,543</td></tr>"
+            "<tr><td>Cash</td><td>$3,845,200</td><td>$6,120,400</td></tr>"
+            "<tr><td>Digital assets</td><td>$352,940</td><td>$1,012,543</td></tr>"
             "</table>"
         ),
     )
@@ -150,21 +150,21 @@ def _mock_direct_lookup_response(chunk_ids: list[str]) -> DirectLookupResponse:
 
 def _mock_table_lookup_response(chunk_id: str) -> DirectLookupResponse:
     """Canned response for table_lookup — answers verbatim from the test
-    table chunk's `table_html` ($4,431,280 and $7,351,900)."""
+    table chunk's `table_html` ($3,845,200 and $6,120,400)."""
     return DirectLookupResponse(
         questions=[
             GeneratedQA(
-                verbatim_span="$4,431,280",
+                verbatim_span="$3,845,200",
                 question="What was cash as of December 31, 2023?",
-                answer="$4,431,280",
+                answer="$3,845,200",
                 answer_type="numeric",
                 difficulty="easy",
                 evidence_chunk_id=chunk_id,
             ),
             GeneratedQA(
-                verbatim_span="$7,351,900",
+                verbatim_span="$6,120,400",
                 question="What was cash as of December 31, 2024?",
-                answer="$7,351,900",
+                answer="$6,120,400",
                 answer_type="numeric",
                 difficulty="easy",
                 evidence_chunk_id=chunk_id,
@@ -634,24 +634,24 @@ def mock_invoke(prompt: str) -> Any:
 
 def test_answer_is_grounded_numeric_exact_match() -> None:
     """Verbatim numeric answer is grounded."""
-    chunk = "Net cash provided by operating activities was $12,945,449 in 2023."
-    assert _answer_is_grounded("$12,945,449", chunk, "numeric") is True
+    chunk = "Net cash provided by operating activities was $14,203,880 in 2023."
+    assert _answer_is_grounded("$14,203,880", chunk, "numeric") is True
 
 
 def test_answer_is_grounded_numeric_rejects_magnitude_rounding() -> None:
     """The whole point of the strict check: '$13 million' is NOT grounded
-    in a chunk that says '$12,945,449' — they differ by far more than the
+    in a chunk that says '$14,203,880' — they differ by far more than the
     0.5 % tolerance the validator uses."""
-    chunk = "Net cash provided by operating activities was $12,945,449 in 2023."
+    chunk = "Net cash provided by operating activities was $14,203,880 in 2023."
     assert _answer_is_grounded("$13 million", chunk, "numeric") is False
     assert _answer_is_grounded("$1 million", chunk, "numeric") is False
 
 
 def test_answer_is_grounded_numeric_tolerates_minor_formatting() -> None:
     """Tiny formatting differences (within 0.5 %) still count as grounded."""
-    chunk = "Cash 7,351,900 at end of year"
-    # 7,351,900 → answer of 7351900 (no commas) should still match
-    assert _answer_is_grounded("7351900", chunk, "numeric") is True
+    chunk = "Cash 6,120,400 at end of year"
+    # 6,120,400 → answer of 6120400 (no commas) should still match
+    assert _answer_is_grounded("6120400", chunk, "numeric") is True
 
 
 def test_answer_is_grounded_string_substring() -> None:
@@ -681,7 +681,7 @@ def test_validator_rejects_ungrounded_numeric_answers(
         "doc_id": "test_doc",
         "chunk_type": "text",
         "chunk_strategy": "section_table",
-        "content": "Net cash provided by operating activities was $12,945,449 in 2023. " * 3,
+        "content": "Net cash provided by operating activities was $14,203,880 in 2023. " * 3,
         "heading_path": ["Cash flow"],
         "child_chunk_ids": [],
         "source_element_ids": ["e1"],
@@ -698,9 +698,9 @@ def test_validator_rejects_ungrounded_numeric_answers(
                 evidence_chunk_id="test_doc:chunk:1",
             ),
             GeneratedQA(
-                verbatim_span="$12,945,449",  # verbatim — must keep
+                verbatim_span="$14,203,880",  # verbatim — must keep
                 question="Net cash from operations in 2023, exact?",
-                answer="$12,945,449",
+                answer="$14,203,880",
                 answer_type="numeric",
                 difficulty="easy",
                 evidence_chunk_id="test_doc:chunk:1",
@@ -736,7 +736,7 @@ def mock_invoke(prompt: str) -> Any:
 
     direct = [r for r in results if r["metadata"]["category"] == "direct_lookup"]
     assert len(direct) == 1
-    assert direct[0]["answer"] == "$12,945,449"
+    assert direct[0]["answer"] == "$14,203,880"
     # No needs_review status should leak into the output
     assert all(r["metadata"].get("validation_status") != "needs_review" for r in results)
 
@@ -956,12 +956,12 @@ def test_entity_audit_rejects_question_containing_corpus_entity() -> None:
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai Inc. is the principal operating entity incorporated in Nevada.",
+            content="Contoso Inc. is the principal operating entity incorporated in Nevada.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
-    # Question that echoes the corpus entity "Neksai Inc." must fail audit
-    assert _entity_audit_fails("How is Neksai Inc. structured?", chunks) is True
+    # Question that echoes the corpus entity "Contoso Inc." must fail audit
+    assert _entity_audit_fails("How is Contoso Inc. structured?", chunks) is True
 
 
 def test_entity_audit_passes_generic_navigational_question() -> None:
@@ -969,7 +969,7 @@ def test_entity_audit_passes_generic_navigational_question() -> None:
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai Inc. is the principal operating entity.",
+            content="Contoso Inc. is the principal operating entity.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
@@ -1002,31 +1002,31 @@ def test_entity_audit_is_case_insensitive() -> None:
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai BVI Ltd. was incorporated in the British Virgin Islands.",
+            content="Contoso BVI Ltd. was incorporated in the British Virgin Islands.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
-    assert _entity_audit_fails("What is neksai bvi ltd.?", chunks) is True
+    assert _entity_audit_fails("What is contoso bvi ltd.?", chunks) is True
 
 
 def test_entity_audit_rejects_single_word_entity() -> None:
-    """First token of a multi-word entity (e.g. 'Neksai' from 'Neksai Inc.') is also audited."""
+    """First token of a multi-word entity (e.g. 'Contoso' from 'Contoso Inc.') is also audited."""
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai Inc. is the principal operating entity.",
+            content="Contoso Inc. is the principal operating entity.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
-    # "Neksai" alone — not the full "Neksai Inc." phrase — must still be rejected.
-    assert _entity_audit_fails("What does Neksai do?", chunks) is True
+    # "Contoso" alone — not the full "Contoso Inc." phrase — must still be rejected.
+    assert _entity_audit_fails("What does Contoso do?", chunks) is True
 
 
 def test_entity_audit_scans_extra_text_fields() -> None:
     """Entity in table_text / pre_text / post_text is also caught."""
     chunk = _make_chunk("doc:chunk:1", content="")
-    chunk["table_text"] = "Neksai BVI Ltd. is a wholly-owned subsidiary."
-    assert _entity_audit_fails("What does Neksai BVI Ltd. do?", [chunk]) is True
+    chunk["table_text"] = "Contoso BVI Ltd. is a wholly-owned subsidiary."
+    assert _entity_audit_fails("What does Contoso BVI Ltd. do?", [chunk]) is True
 
 
 def test_primary_leaf_is_deterministic() -> None:
@@ -1058,7 +1058,7 @@ def test_navigational_prompt_omits_chunk_content(mock_settings: MagicMock) -> No
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="SECRET_CONTENT: Neksai Inc. revenue was $1M",
+            content="SECRET_CONTENT: Contoso Inc. revenue was $1M",
             heading_path=["CORPORATE HISTORY AND STRUCTURE"],
         ),
         _make_chunk(
@@ -1135,7 +1135,7 @@ def test_navigational_example_schema(mock_settings: MagicMock) -> None:
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="The company operates through Neksai Inc. and two BVI subsidiaries.",
+            content="The company operates through Contoso Inc. and two BVI subsidiaries.",
             heading_path=["CORPORATE HISTORY AND STRUCTURE"],
         ),
     ]
@@ -1179,7 +1179,7 @@ def test_navigational_entity_leak_rejected_in_run(mock_settings: MagicMock) -> N
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai Inc. operates the principal business.",
+            content="Contoso Inc. operates the principal business.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
@@ -1188,7 +1188,7 @@ def mock_invoke(prompt: str) -> NavigationalResponse:
         return NavigationalResponse(
             questions=[
                 NavigationalQA(
-                    question="What does Neksai Inc. do?",  # entity leak — must be dropped
+                    question="What does Contoso Inc. do?",  # entity leak — must be dropped
                     target_section="CORPORATE STRUCTURE",
                     representative_answer="It operates the principal business.",
                 ),
@@ -1212,7 +1212,7 @@ def mock_invoke(prompt: str) -> NavigationalResponse:
 
     nav = [r for r in results if r["metadata"]["category"] == "navigational"]
     assert len(nav) == 1
-    assert "Neksai" not in nav[0]["question"]
+    assert "Contoso" not in nav[0]["question"]
 
 
 @patch("tools.dataset_generation.llm.get_settings")
@@ -1223,7 +1223,7 @@ def test_navigational_run_tops_up_after_entity_rejections(mock_settings: MagicMo
     chunks = [
         _make_chunk(
             "doc:chunk:1",
-            content="Neksai Inc. operates the principal business.",
+            content="Contoso Inc. operates the principal business.",
             heading_path=["CORPORATE STRUCTURE"],
         ),
     ]
@@ -1233,7 +1233,7 @@ def test_navigational_run_tops_up_after_entity_rejections(mock_settings: MagicMo
             NavigationalResponse(
                 questions=[
                     NavigationalQA(
-                        question="What does Neksai Inc. do?",
+                        question="What does Contoso Inc. do?",
                         target_section="CORPORATE STRUCTURE",
                         representative_answer="It operates the principal business.",
                     ),
diff --git a/tests/unit/test_parser_preprocess.py b/tests/unit/test_parser_preprocess.py
index 717210a..92b15d2 100644
--- a/tests/unit/test_parser_preprocess.py
+++ b/tests/unit/test_parser_preprocess.py
@@ -32,7 +32,7 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
                 "avatar_comments_doc": [
                     {
                         "type": "UncategorizedText",
-                        "text": "NEKSAI INC.",
+                        "text": "CONTOSO INC.",
                         "element_id": "boilerplate-1",
                         "metadata": {
                             "category_depth": 0,
@@ -42,7 +42,7 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
                     },
                     {
                         "type": "UncategorizedText",
-                        "text": "NEKSAI INC.",
+                        "text": "CONTOSO INC.",
                         "element_id": "boilerplate-2",
                         "metadata": {
                             "category_depth": 0,
@@ -52,7 +52,7 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
                     },
                     {
                         "type": "UncategorizedText",
-                        "text": "NEKSAI INC.",
+                        "text": "CONTOSO INC.",
                         "element_id": "boilerplate-3",
                         "metadata": {
                             "category_depth": 0,
@@ -96,7 +96,7 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
                     },
                     {
                         "type": "Table",
-                        "text": "Cash $ 4,431,280 $ 7,351,900",
+                        "text": "Cash $ 3,845,200 $ 6,120,400",
                         "element_id": "table-1",
                         "metadata": {
                             "filename": "avatar-with-comments.docx",
@@ -104,7 +104,7 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
                                 "<table>"
                                 "<tr><td></td><td>December 31, 2023</td><td>December 31, 2023</td>"
                                 "<td></td><td>December 31, 2024</td><td>December 31, 2024</td></tr>"
-                                "<tr><td>Cash</td><td>$</td><td>4,431,280</td><td>$</td><td>7,351,900</td><td></td></tr>"
+                                "<tr><td>Cash</td><td>$</td><td>3,845,200</td><td>$</td><td>6,120,400</td><td></td></tr>"
                                 "</table>"
                             ),
                         },
@@ -138,17 +138,17 @@ def test_run_parser_preprocess_pipeline_normalizes_elements(tmp_path: Path) -> N
     table_element = next(
         element for element in normalized_elements if element["element_id"] == "table-1"
     )
-    assert "$4,431,280" in table_element["text"]
+    assert "$3,845,200" in table_element["text"]
     assert "December 31, 2023" in table_element["text"]
     assert "December 31, 2024" in table_element["text"]
 
     review_html = Path(result.review_html_path).read_text()
     rendered_rows = _table_rows_from_html(review_html)
-    assert "$4,431,280" in review_html
-    assert "$7,351,900" in review_html
+    assert "$3,845,200" in review_html
+    assert "$6,120,400" in review_html
     assert "><td>$</td>" not in review_html
     assert rendered_rows[0] == ["", "December 31, 2023", "", "December 31, 2024"]
-    assert rendered_rows[1] == ["Cash", "", "$4,431,280", "$7,351,900"]
+    assert rendered_rows[1] == ["Cash", "", "$3,845,200", "$6,120,400"]
 
     warning_codes = {warning.code for warning in result.warnings}
     assert "missing_page_numbers" in warning_codes
@@ -214,14 +214,14 @@ def test_run_parser_preprocess_pipeline_can_enable_table_column_merge(tmp_path:
             [
                 {
                     "type": "Table",
-                    "text": "Cash $ 4,431,280 $ 7,351,900",
+                    "text": "Cash $ 3,845,200 $ 6,120,400",
                     "element_id": "table-1",
                     "metadata": {
                         "text_as_html": (
                             "<table>"
                             "<tr><td></td><td>December 31, 2023</td><td>December 31, 2023</td>"
                             "<td></td><td>December 31, 2024</td><td>December 31, 2024</td></tr>"
-                            "<tr><td>Cash</td><td>$</td><td>4,431,280</td><td>$</td><td>7,351,900</td><td></td></tr>"
+                            "<tr><td>Cash</td><td>$</td><td>3,845,200</td><td>$</td><td>6,120,400</td><td></td></tr>"
                             "</table>"
                         ),
                     },
@@ -242,7 +242,7 @@ def test_run_parser_preprocess_pipeline_can_enable_table_column_merge(tmp_path:
     review_html = Path(result.review_html_path).read_text()
     rendered_rows = _table_rows_from_html(review_html)
     assert rendered_rows[0] == ["", "December 31, 2023", "December 31, 2024"]
-    assert rendered_rows[1] == ["Cash", "$4,431,280", "$7,351,900"]
+    assert rendered_rows[1] == ["Cash", "$3,845,200", "$6,120,400"]
 
 
 def test_run_parser_preprocess_pipeline_describes_images_when_enabled(tmp_path: Path) -> None:
@@ -441,14 +441,14 @@ def test_parser_preprocess_command_accepts_config(tmp_path: Path) -> None:
             [
                 {
                     "type": "Table",
-                    "text": "Cash $ 4,431,280 $ 7,351,900",
+                    "text": "Cash $ 3,845,200 $ 6,120,400",
                     "element_id": "table-1",
                     "metadata": {
                         "text_as_html": (
                             "<table>"
                             "<tr><td></td><td>December 31, 2023</td><td>December 31, 2023</td>"
                             "<td></td><td>December 31, 2024</td><td>December 31, 2024</td></tr>"
-                            "<tr><td>Cash</td><td>$</td><td>4,431,280</td><td>$</td><td>7,351,900</td><td></td></tr>"
+                            "<tr><td>Cash</td><td>$</td><td>3,845,200</td><td>$</td><td>6,120,400</td><td></td></tr>"
                             "</table>"
                         ),
                     },
diff --git a/tests/unit/test_verify_executors.py b/tests/unit/test_verify_executors.py
index 3c8d5da..11ee486 100644
--- a/tests/unit/test_verify_executors.py
+++ b/tests/unit/test_verify_executors.py
@@ -25,7 +25,7 @@ def _write_base(tmp_path: Path) -> Path:
                     "doc_ids": ["f1-doc", "pj-doc"],
                 },
                 "sources": [
-                    {"path": "data/raw/ir_document/F1-file.docx", "doc_id": "f1-doc"},
+                    {"path": "data/raw/ir_document/ir-sample-b.docx", "doc_id": "f1-doc"},
                     {"path": "data/raw/ir_document/Pj.docx", "doc_id": "pj-doc"},
                 ],
                 "vector_store": {
@@ -62,7 +62,7 @@ def test_derive_config_rewrites_pipeline_and_paths(tmp_path: Path) -> None:
     assert derived["pipeline"]["artifact_dir"] == str(artifact_dir)
     # Sources keep their doc_id, paths rewritten under raw_dir by basename.
     assert derived["sources"] == [
-        {"path": str(raw_dir / "F1-file.docx"), "doc_id": "f1-doc"},
+        {"path": str(raw_dir / "ir-sample-b.docx"), "doc_id": "f1-doc"},
         {"path": str(raw_dir / "Pj.docx"), "doc_id": "pj-doc"},
     ]
     assert derived["vector_store"]["options"]["path"] == str(qdrant_path)
diff --git a/tools/dataset_generation/llm.py b/tools/dataset_generation/llm.py
index 7af8c7e..d1708ca 100644
--- a/tools/dataset_generation/llm.py
+++ b/tools/dataset_generation/llm.py
@@ -149,7 +149,7 @@ def _normalize_for_match(text: str) -> str:
 def _extract_numbers(text: str) -> list[float]:
     """Pull all numbers from a string (handles commas, decimals, magnitudes)."""
     out: list[float] = []
-    # Find tokens like 12,945,449 or 5.9 (without magnitude)
+    # Find tokens like 14,203,880 or 5.9 (without magnitude)
     for m in re.finditer(r"-?[\d,]*\d[\d,]*\.?\d*", text or ""):
         token = m.group()
         try:
@@ -166,7 +166,7 @@ def _answer_is_grounded(
 
     For numeric answers: every number in the answer must appear in the chunk
     (within `tol` relative tolerance — default 0.5 % to account for cell
-    formatting differences but reject magnitude rounding like 5.9M vs 5,898,348).
+    formatting differences but reject magnitude rounding like 5.9M vs 6,357,910).
 
     For string answers: the answer's normalized content must be a substring
     of the normalized chunk.
@@ -260,9 +260,9 @@ def _build_corpus_entities(chunks: list[dict[str, Any]]) -> set[str]:
     """Extract a lowercase entity set from all text fields across chunks.
 
     Scans ``content``, ``table_text``, ``pre_text``, and ``post_text``.
-    Multi-word entity phrases (e.g. ``"neksai inc."``) are stored whole; the
+    Multi-word entity phrases (e.g. ``"contoso inc."``) are stored whole; the
     first token is also added when it has ≥5 characters and is not a known legal
-    suffix so that single-word usage (e.g. ``"Neksai"``) is caught by the audit.
+    suffix so that single-word usage (e.g. ``"Contoso"``) is caught by the audit.
 
     Args:
         chunks: Chunk dicts from chunks.json.
@@ -667,20 +667,20 @@ def _generate_table_lookup(
             "3. Copy `verbatim_span` unchanged into `answer`.\n\n"
             "**Strict formatting rules (violations cause the row to be "
             "rejected):**\n"
-            '- NO magnitude compression: if the cell shows "$12,945,449" '
-            'the verbatim_span must be "$12,945,449", NOT "$13M" or '
+            '- NO magnitude compression: if the cell shows "$14,203,880" '
+            'the verbatim_span must be "$14,203,880", NOT "$13M" or '
             '"about $13 million".\n'
             "- NO unit conversion: copy the number at the scale it appears.\n"
             '- NO paraphrasing percentages: copy "33.7%" not "~34%".\n'
             "- Preserve the source's sign convention (parentheses for "
             "negatives if the table uses them).\n"
             "- Do NOT append unit words to numbers "
-            '("$7,351,900" not "$7,351,900 dollars").\n\n'
+            '("$6,120,400" not "$6,120,400 dollars").\n\n'
             "**Few-shot examples:**\n\n"
             "GOOD:\n"
-            '  verbatim_span: "$4,431,280"\n'
+            '  verbatim_span: "$3,845,200"\n'
             '  question: "What was the cash balance as of December 31, 2023?"\n'
-            '  answer: "$4,431,280"\n'
+            '  answer: "$3,845,200"\n'
             "  answer_type: numeric\n\n"
             "GOOD:\n"
             '  verbatim_span: "33.7%"\n'
@@ -689,9 +689,9 @@ def _generate_table_lookup(
             '  answer: "33.7%"\n'
             "  answer_type: numeric\n\n"
             "BAD — never do this:\n"
-            '  verbatim_span: "about $4.4 million"   ← compressed; not in '
+            '  verbatim_span: "about $3.8 million"   ← compressed; not in '
             "the table\n"
-            '  answer: "$4.4M"   ← also wrong\n\n'
+            '  answer: "$3.8M"   ← also wrong\n\n'
             "Other rules:\n"
             "- Each question must be answerable from exactly ONE table chunk.\n"
             "- The question MUST name the exact time period from the column "
@@ -818,20 +818,20 @@ def _call_direct_lookup(self, chunks: list[dict[str, Any]], count: int) -> list[
             "If you cannot find a suitable verbatim span, skip that fact and "
             "pick a different one.\n\n"
             "**Strict formatting rules (violations cause rejection):**\n"
-            '- NO magnitude compression: "$12,945,449" not "$13M" or '
+            '- NO magnitude compression: "$14,203,880" not "$13M" or '
             '"~$13 million".\n'
             "- NO paraphrasing: copy names, dates, percentages exactly as "
             "written in the chunk.\n"
-            '- NO unit words appended to numbers: "$7,351,900" not '
-            '"$7,351,900 dollars".\n'
+            '- NO unit words appended to numbers: "$6,120,400" not '
+            '"$6,120,400 dollars".\n'
             "- For string answers, use the shortest verbatim span — no "
             "surrounding narrative.\n\n"
             "**Few-shot examples:**\n\n"
             "GOOD:\n"
-            '  verbatim_span: "$12,945,449"\n'
+            '  verbatim_span: "$14,203,880"\n'
             '  question: "What was net cash provided by operating activities'
             ' for the year ended December 31, 2023?"\n'
-            '  answer: "$12,945,449"\n'
+            '  answer: "$14,203,880"\n'
             "  answer_type: numeric\n\n"
             "GOOD:\n"
             '  verbatim_span: "AI Avatar Co., Ltd"\n'
diff --git a/tools/dataset_generation/template.py b/tools/dataset_generation/template.py
index 7b6fc76..d9079e0 100644
--- a/tools/dataset_generation/template.py
+++ b/tools/dataset_generation/template.py
@@ -216,7 +216,7 @@ def _parse_numeric(text: str) -> float | None:
     """Parse a financial numeric string into a float.
 
     Args:
-        text: Raw cell value like "$4,431,280" or "(317,662)".
+        text: Raw cell value like "$3,845,200" or "(352,940)".
 
     Returns:
         Parsed float value, or None if unparseable.
diff --git a/xrag/core/generation/openai.py b/xrag/core/generation/openai.py
index 54c6e99..0f23d4a 100644
--- a/xrag/core/generation/openai.py
+++ b/xrag/core/generation/openai.py
@@ -19,7 +19,7 @@
     "Numeric extraction rules — when the answer is a number from a table:\n"
     "  1. Return the EXACT value as it appears in the source, including "
     "every digit and the original formatting "
-    '(e.g. "$12,945,449", not "$13M" or "~$13 million"). Do NOT '
+    '(e.g. "$14,203,880", not "$13M" or "~$13 million"). Do NOT '
     "round, approximate, or compress to a magnitude word.\n"
     "  2. If multiple cells could match, pick the ONE whose row label "
     "AND column header EXACTLY match the question's metric name and "
@@ -40,8 +40,8 @@
     "#3 contains it. When in doubt, attempt to extract.\n"
     "\n"
     "When a calculation is required, show the component values from "
-    "the context first (e.g., 'Revenue 2024 = $45,042,408, Revenue "
-    "2023 = $23,761,586'), then state the computed result. Never "
+    "the context first (e.g., 'Revenue 2024 = $48,715,220, Revenue "
+    "2023 = $26,118,340'), then state the computed result. Never "
     "invent values not present in the context. Be concise."
 )
 
@@ -78,8 +78,8 @@ def _chunks_to_documents(chunks: list[dict[str, Any]]) -> list[Document]:
     back to markdown.
 
     TODO: improve ``src/core/common/converter.py:table_html_to_markdown``
-    to drop empty pad columns, merge currency cells (``$ | 4,431,280``
-    → ``$4,431,280``), and collapse duplicate multi-row date headers.
+    to drop empty pad columns, merge currency cells (``$ | 3,845,200``
+    → ``$3,845,200``), and collapse duplicate multi-row date headers.
     Then flip the priority above so markdown wins by default and HTML
     is the fallback.
     """