Intent
Improve LLM outputs by injecting indexed knowledge.
When to use
- You need answers grounded in a document set.
- Users ask questions over internal docs.
- You want explainable citations.
Core mechanics
- Ingest documents and chunk text.
- Embed and index chunks.
- Retrieve relevant context and prompt generation.
Implementation checklist
- Define source corpus and update cadence.
- Choose chunking and embedding strategies.
- Build and validate the index.
- Implement retrieval with relevance scoring.
- Log citations and evaluate responses.
Failure modes and mitigations
- Stale index -> schedule refresh and versioning.
- Low recall -> tune chunking and embeddings.
- Hallucinations -> enforce citation requirements.
Observability and validation
- Retrieval hit rate and top sources.
- Evaluation scores and user feedback.
Artifacts
- Index build logs.
- Evaluation dataset and results.