Knowledge retrieval troubleshooting for agents
Key Takeaway
Find the layer that failed before you fix anything. When part of a document is missing from answers, increase the number of results the retriever returns. When archived Knowledge articles cause outdated answers, add a retriever prefilter so that only active articles are returned.
An agent grounded with retrieval augmented generation (RAG) can give missing, wrong, or outdated answers. The cause can sit in the agent's reasoning, the knowledge content, the Data 360 search index and retriever, or the LLM's generated answer. Troubleshooting means finding the layer that failed, then fixing it with tools such as Retriever Playground, citations, Enhanced Event Logs, and RAG quality metrics. Poor content quality is the most common cause.
How to troubleshoot layer by layer
Work down the stack one layer at a time, starting with the agent.
- Check the agent layer. In Agentforce Builder or Agentforce Testing Center, check whether the Atlas reasoning engine selects the right subagent and runs the right action. If it doesn't, the problem is most likely in the subagent instructions, action instructions, and classification descriptions.
- If the agent layer is fine and the solution uses an Agentforce Data Library, confirm permissions and check the data library status.
- Check the search index and retrieval layer. Find out whether the search index holds the content you expect, and which chunks come back.
- Check the generation layer. Compare the retrieved chunks in the prompt with the LLM's response to find what is incorrect, missing, or hallucinated.
If you fix the problem at the agent layer, you don't need to troubleshoot the RAG components.