Rag And Grounded Generation

Operator supplies retrieved passages to /v2/chat as documents[] and wants the answer grounded only in them, but provides each document as a bare string with no id.

Provide each document with a stable id and structured data so the response can cite document ids precisely. Stable ids let the integrator map citations back to source records; bare strings make citation attribution ambiguous.

User asks a question whose answer is absent from the supplied documents[]. The model is expected to ground its answer in those documents.

When the documents do not support an answer, the grounded response should abstain or state that the documents do not contain the answer, rather than fabricate. The integrator should treat an uncited confident claim as a grounding failure and route it for review.

A retrieval pipeline passes the top 100 vector hits straight into /v2/chat documents[] without reranking, blowing the context budget and diluting grounding.

Rerank candidate passages (e.g., via /v2/rerank) and pass only the top_n most relevant into documents[] for grounded generation. Tighter, higher-precision context improves citation quality and reduces token cost.