RAG retrieval sources are passed through to the model context

A RAG pipeline that retrieves documents but fails to inject them into the model context — or injects them as unstructured blob text — is just an expensive vector search that the LLM cannot use. The model falls back to parametric knowledge, hallucinates confidently on domain-specific questions, and cites sources it never saw. This breaks the inference-contract taxon: you promised grounded answers, you delivered confabulation. Users trust the retrieved citations exactly when they are least reliable.

Wrap each retrieved chunk in labeled delimiters before injection, and instruct the model to answer only from the provided context and acknowledge insufficient context explicitly. Apply this at the retrieval boundary in lib/rag.ts.

const contextBlock = docs.map((d, i) => `<source id="${i+1}" title="${d.metadata.title}">\n${d.pageContent}\n</source>`).join('\n')
const userMessage = `Answer from the sources below. If insufficient, say so.\n\n${contextBlock}\n\nQuestion: ${q}`