RAG and document ingestion checklist
From source documents to cited answers.
DeveloperContent ownerQA
Write the document
- Descriptive title, audience and region
- Metadata block: id, version, status, effective date, owner
- One topic per section; the answer first
- Scope stated where ambiguous
- Real text, tables as Markdown or CSV, alt text for images
Manage the corpus
- One source of truth per topic
- Versions, validity windows and supersession
- Filter retrieval to current documents
- Enforce access control at retrieval time
- Parallel documents per language
Pipeline
- Extract, clean, enrich with metadata
- Chunk by structure with context headers
- Stable chunk IDs; hash sources; re-ingest only changes
- Hybrid search (keyword plus semantic) with reranking
- Relevance threshold and a not-found path
Test before launch
- Golden questions with expected passages
- Retrieve-then-read check
- Conflict and freshness scans
- Access and not-found tests