RAG & Knowledge Assistants

RAG supplies retrieved material to a model at request time. Define which documents are in scope, who can access them and how freshness will be maintained.

Clean extraction problems before splitting documents. Preserve headings, source identifiers and useful metadata. Chunk boundaries should keep meaning together rather than merely reach a fixed size.

Embeddings support similarity search; keyword filters and metadata can add useful constraints. Inspect retrieved passages before judging the generated answer.

Pass retrieved passages as untrusted reference material. Ask the model to cite available source identifiers and to say when the provided material is insufficient. Validate references against the retrieved set.

Separate retrieval failures from generation failures. Include answerable questions, ambiguous questions and questions outside the collection. Change one retrieval choice at a time.