Skip to main content
Retrieval-augmented generation, or RAG, is intended to bring relevant source material into an agent’s response. Prepare clear documents and a small set of questions with known answers before adjusting retrieval settings.
The API exposes index, chunk, and query operations. An index marked successful or a configured embedding-model identifier does not establish semantic vector retrieval or a live knowledge-grounded answer. Conversation execution remains gated.

​
Prepare and inspect the source

1

Read the stored document

Follow document management. Check headings, current facts, and source ownership before creating an index. Empty or incomplete source content cannot provide useful answers.
2

Review index state and chunks

Use Get document RAG indexes to inspect the index state. Create document RAG index requires a model selection and write access.Inspect document chunks. Check that important instructions and their qualifications remain understandable. Do not interpret the index status as proof that an embedding provider ran.
3

Review the agent attachment

Attach the source to an existing agent through conversation_config.agent.prompt.knowledge_base. Each locator uses type, name, id, and usage_mode, whose values are auto or prompt. Initial agent creation rejects nonempty knowledge attachments.
4

Prepare a query review

The Query agent knowledge base reference describes a separate query operation. Review any returned material against the known answer and source text. A useful result from this operation does not demonstrate retrieval inside a live conversation.

​
Retrieval configuration

These fields belong to conversation_config.agent.prompt.rag. Save them in a complete server draft and inspect the read-back. These are authoring limits. A specific query operation may impose narrower limits, and a schema field does not establish that every retrieval stage uses it. Avoid tuning values until you can inspect the actual returned material.

​
Frequently asked questions

No. A configuration flag does not prove retrieval, answer grounding, or citation behavior. Keep those as separate checks in the eventual conversation review.
Include an exact factual question, a paraphrase, a question spanning two sections, and a question the source cannot answer. Record which passage supports each expected answer.