Skip to main content

Knowledge Platform

The Knowledge Platform indexes organizational content for retrieval-augmented generation (RAG). Uploads and crawls are chunked, embedded, and retrieved with hybrid BM25 + vector search. Answers include citations; the assistant is guided to refuse when no relevant source exists.

Introduction​

Ingest paths in the real API:

  • POST /api/v1/documents — multipart upload (quota: check_documents_limit)
  • POST /api/v1/documents/text — create from raw text
  • POST /api/v1/documents/:id/reindex — rebuild index for a document
  • GraphQL mutations for console workflows including website crawl
  • GET /api/v1/documents/:id/view — authorized document view

Supported formats include PDF, DOCX, Markdown, TXT; OCR covers scans/images. Multilingual retrieval supports EN, AR, TA, HI and more.

Why it exists​

Customer AI and Employee AI must share one high-quality retrieval stack with hard workspace isolation.

Concepts​

  • Document / chunk / embedding — indexed units
  • Hybrid retrieval — lexical + vector
  • Citation — source attribution in answers
  • Isolation — workspace-scoped indexes

Architecture​

Workflow​

Improve answer quality

  1. Curate sources — Prefer canonical policies over stale copies.
  2. Ingest — Upload or crawl; wait for indexing.
  3. Probe — Ask known questions; verify citations.
  4. Reindex — Use reindex after content corrections.

Code examples​

curl -sS -X POST \
  -H "Authorization: Bearer $USER_JWT" \
  -F "[email protected]" \
  -F "workspace_id=$WORKSPACE_ID" \
  https://api.qefro.com/api/v1/documents

Best practices​

  • Separate customer vs employee corpora by workspace
  • Re-test after every major document delete/replace
  • Prefer crawl allowlists over unbounded domains

Security notes​

FAQ​

Do I host the vector DB?
Not on Qefro cloud — indexing and retrieval are platform-managed.