.env.example

# Copy to .env and fill in. .env is gitignored and must never be committed.
#
# Every value here is a placeholder. The pipeline runs with none of them set —
# `python pipelines/run_local_demo.py` uses the filesystem and the in-process
# index. You only need these to point it at a real workspace.

# ---------------------------------------------------------------------------
# Workspace credentials — required only for a remote run
# ---------------------------------------------------------------------------
# Read from the environment, never written to logs. Settings.redacted() masks
# them, and Settings has repr=False on both fields.
DATABRICKS_HOST=https://your-workspace.cloud.databricks.example
DATABRICKS_TOKEN=

# ---------------------------------------------------------------------------
# Storage
# ---------------------------------------------------------------------------
# The governed landing zone. On Databricks this is a Unity Catalog Volume path
# such as /Volumes/<catalog>/<schema>/<volume>/<prefix>.
KB_LANDING_ROOT=./data/landing

# Where authored knowledge-base documents are written.
KB_OUTPUT_ROOT=./data/knowledge_base

# Extraction artifacts and run state (manifest, incremental state).
KB_EXTRACTED_ROOT=./data/extracted
KB_STATE_ROOT=./data/state

# ---------------------------------------------------------------------------
# Logical scope
# ---------------------------------------------------------------------------
# Every chunk is stamped with both, and every query is filtered by both. This
# is the isolation boundary: content from one collection cannot be retrieved by
# a query scoped to another.
KB_PROJECT_ID=example_project
KB_COLLECTION_ID=example_collection

# ---------------------------------------------------------------------------
# Warehouse identifiers
# ---------------------------------------------------------------------------
KB_CATALOG=portfolio_catalog
KB_SCHEMA=knowledge_base
KB_VOLUME=raw_documents

# ---------------------------------------------------------------------------
# Model and search endpoints
# ---------------------------------------------------------------------------
# Names of your own serving endpoints. Left unset, the pipeline uses the
# deterministic passthrough client and generates nothing.
KB_LLM_ENDPOINT=
KB_EMBEDDING_ENDPOINT=
KB_VECTOR_ENDPOINT=

# Set true only once KB_LLM_ENDPOINT points at a real endpoint. Settings
# validation rejects the inconsistent combination.
KB_LLM_ENABLED=false

# ---------------------------------------------------------------------------
# Behaviour
# ---------------------------------------------------------------------------
# Compute and report every action, write nothing. Useful before a first run.
KB_DRY_RUN=false

# Reprocess sources whose content hash and extractor version are both unchanged.
KB_FORCE_REPROCESS=false

# ---------------------------------------------------------------------------
# Retrieval
# ---------------------------------------------------------------------------
KB_TOP_K=8

# Leave unset. Score scales differ between backends, so each one declares its
# own calibrated floor and retrieval asks for it. Set this only to override.
KB_SIMILARITY_THRESHOLD=

# Cap on context assembled per query. Retrieval that returns 12k tokens has not
# answered anything, it has moved the problem into the model's context window.
KB_MAX_CONTEXT_TOKENS=2000

# Widen a retrieved chunk to its parent section when the budget allows.
KB_EXPAND_HIERARCHY=true