Feature: Cross-Document Semantic Search & “Ask Your Workspace” (RAG across all uploads)

[hoangsonww]

Summary

Enable semantic search and grounded Q&A across all documents in a user’s library (web, mobile, and VS Code extension). The feature indexes uploaded content into a vector store, supports filters (type, tags, date), and powers an “Ask Your Workspace” chat that returns cited answers with source snippets.

---

Motivation

• Users upload many files; today they must open each to find answers.
• RAG across the user’s corpus unlocks instant insights, better chat grounding, and stronger reuse of prior uploads.
• Pairs well with existing Redis/RabbitMQ infra for indexing and caching; exposes clean GraphQL/REST for all clients (web, mobile, extension).

---

Goals

1. Semantic Search across a user’s entire library with filters (file type, tags, date range).
2. Workspace Q&A: natural-language questions with answers grounded in top-k passages + inline citations.
3. Citations UI: expandable snippets showing exact source, page/section, and quick-open.
4. Indexing Pipeline: chunking, dedup (content hash), incremental updates on new/edited docs.
5. Privacy & Isolation: strict per-user namespaces; no cross-tenant leakage.
6. VS Code Extension integration: global search & “Ask Workspace” panel.
7. Admin/Debug endpoints: reindex status, queue depth, last error.

Non-Goals (this issue): org-wide sharing, external drives sync, full OCR/table extraction (can be follow-ups).

---

Acceptance Criteria

• A global search bar (web) returns relevant results in <1.5s p95 for top-k=10 on a typical user library (≤ 500 docs).
• Clicking a result opens the doc page scrolled to the cited passage (or shows an inline snippet).
• “Ask Your Workspace” answers stream tokens and include at least 3 citations when available; if no sources found, return a clear “no sources” state (no hallucinated content).
• Filters: MIME (pdf/docx/txt), tags, created/updated date range.
• Works in dark/light modes; fully keyboard accessible (enter to search, arrows to navigate results).
• Extension exposes the same search & Q&A (no extra login beyond current app auth).
• Telemetry events emitted (see Analytics).

---

UX Notes

• Global Search in header; results list with snippet, confidence, tags, and file icon.
• Q&A Panel (ChatModal): a toggle “Search scope: current doc | all docs”. Answers include numbered citations [¹][²][³]; clicking jumps to snippet.
• Empty State: suggests uploading docs or widening filters.

---

API & GraphQL (proposed)

REST (Express):

POST /v1/search/semantic
body: { query: string, topK?: number, filters?: { mimeTypes?: string[], tags?: string[], dateFrom?: string, dateTo?: string } }

POST /v1/qa/workspace
body: { question: string, topK?: number, filters?: {...}, stream?: boolean }

GET  /v1/index/status
POST /v1/index/reindex   # optional: specific docIds or full user reindex

GraphQL:

type SearchHit {
  docId: ID!
  title: String!
  snippet: String!
  score: Float!
  location: String    # page/section/offset
  mimeType: String
  tags: [String!]
  updatedAt: String
}

type QAChunk {
  text: String!
  citations: [SearchHit!]!
}

type Query {
  workspaceSearch(query: String!, topK: Int, filters: FiltersInput): [SearchHit!]!
  workspaceQA(question: String!, topK: Int, filters: FiltersInput): [QAChunk!]!
}

input FiltersInput {
  mimeTypes: [String!]
  tags: [String!]
  dateFrom: String
  dateTo: String
}

Response example:

{
  "answer": "Project X uses a zero-downtime rollout via NGINX and Kubernetes.",
  "citations": [
    { "docId": "abc123", "title": "kubernetes/README.md", "snippet": "NGINX as reverse proxy...", "location": "L45-L67" }
  ]
}

---

Indexing Pipeline

• Chunking: 800–1200 tokens, 150–200 overlap; include docId, page/section, content hash.
• Embeddings: pluggable provider; store dim in config (e.g., 1536).
• Vector Store: start with FAISS/pgvector (self-host) and adapter pattern for Pinecone/Weaviate later.
• Ingestion: PDFs/Docx → text; reuse existing upload flow; enqueue index_document via RabbitMQ.
• Dedup: skip chunks with unchanged hash.
• Metadata: MongoDB holds doc + chunk metadata; Redis caches top-k for frequent queries.

---

Security & Privacy

• Verify Firebase/JWT at gateway; namespace vectors by userId.
• Encrypt at rest when supported; redact secrets/PII heuristically (optional flag).
• Rate-limit search/QA to prevent abuse.

---

Performance

• Async indexing; UI shows per-doc index status.
• Redis cache for popular queries (TTL 15–60m).
• Stream answers over SSE/WebSocket.
• Observability: log recall@k on sampled queries to track quality.

---

Analytics / Events

• workspace_search {queryLen, topK, filters, hits}
• workspace_qa {questionLen, citedCount, latencyMs}
• citation_click {docId, location}
• KPIs: time-to-first-answer, retention uplift for users who use search/QA, zero-result rate.

---

Rollout Plan

1. Phase 1: Indexing + semantic search (web) with filters and citations.
2. Phase 2: Workspace Q&A + streaming + VS Code extension integration.
3. Phase 3: Mobile support + admin tools (reindex UI, queue health).

---

Tasks

Backend (Express)

• Add /v1/search/semantic, /v1/qa/workspace, /v1/index/* routes + validators.
• Implement chunking+embeddings service; vector store adapter; Redis caching.
• Queue workers (RabbitMQ) for indexing; idempotent updates by content hash.
• Unit/integration tests; load tests for p95 latency.

GraphQL

• Schema types + resolvers for workspaceSearch and workspaceQA.
• Auth directives to enforce per-user isolation.

Frontend (React)

• Global search bar + results list with filters drawer.
• Update ChatModal: scope toggle + citation rendering + copy/export.
• Empty/Loading/Skeleton states; a11y + keyboard nav.

VS Code Extension

• Add “Ask Workspace” view + semantic search command.
• Reuse existing auth; streaming UI.

Mobile (Expo)

• Minimal semantic search screen + deep-link to document.
• Optional: offline cache of last queries.

Infra/CI

• Feature flag: workspace_search_enabled.
• Dashboards: queue depth, index lag, query latency, error rate.

Docs

• Update README.md + openapi.yaml + extension docs.
• Add “Quality & Evaluation” section (recall@k, groundedness).

---

Open Questions

• Preferred default vector backend? (pgvector vs FAISS vs managed)
• Maximum library size before tiered indexing needed?
• OCR/table extraction scope now vs. follow-up issue?
• How should tags be managed (freeform vs predefined) for better filters?

---