Retrievers And Query Engines

A RetrieverQueryEngine is left at similarity_top_k=2 for a corpus where relevant evidence is routinely spread across 5-8 chunks; answers are incomplete.

Tune similarity_top_k to the evidence-spread of the corpus and the LLM's context budget — raise k (and optionally add a reranker to trim) when answers need more chunks, while watching context overflow and cost. Treat k as a measured parameter, not a default.

A query engine returns a Response. The application renders response.response text to users as cited answers but never inspects response.source_nodes.

Use response.source_nodes (NodeWithScore) to attribute each answer to the retrieved nodes — render citations from node metadata (source, page) and verify the answer's claims are supported by those nodes. An answer with no/low-score source nodes should be flagged as ungrounded, not presented as cite…

Compliance requires inline source markers in answers. The integrator uses a plain RetrieverQueryEngine and post-hoc asks the LLM to 'add citations', which fabricates source numbers.

Use CitationQueryEngine (or equivalent) which splits sources into citation chunks and prompts the model to cite by source number tied to actual retrieved nodes, so [N] markers map to real source_nodes. Verify each emitted citation maps to an existing retrieved chunk, not an invented index.