In the knowledge retrieval step of ChatFlow, configuring two knowledge bases fails to retrieve any results.

[luboyan6]

Self Checks

• I have searched for existing issues search for existing issues, including closed ones.
• I confirm that I am using English to submit this report (我已阅读并同意 Language Policy).
• [FOR CHINESE USERS] 请务必使用英文提交 Issue，否则会被关闭。谢谢！:)
• Please do not modify this template :) and fill in all the required fields.

Content

In the knowledge retrieval step of ChatFlow, when two knowledge bases are configured, no results can be retrieved; however, if either one of them is configured individually, knowledge retrieval works as expected. How can we analyze and troubleshoot this issue, or alternatively, work around it using other approaches?

[dosubot[bot]]

When retrieval works with one knowledge base but fails with two, it's usually due to configuration mismatches or constraints in how Dify merges results from multiple sources. Here are the most common causes and steps to troubleshoot:

• If your knowledge bases use different indexing techniques or embedding models, and reranking is not enabled (or not set to use a reranking model), retrieval will fail with a ValueError. Make sure all knowledge bases use the same indexing technique and embedding model, or enable reranking with a reranking model in the retrieval node settings. This is a documented requirement in the codebase and has been reported in community discussions (source).

• Metadata filtering (disabled, automatic, or manual) can restrict which documents are considered. If filters are too restrictive or misconfigured, you may get empty results when multiple knowledge bases are active. Try disabling metadata filtering to rule this out (source).

• Double-check that the knowledge retrieval node in ChatFlow is explicitly linked to the correct datasets, and that retrieval parameters (search mode, rerank model, top_k, score threshold) match those used in your successful single-knowledge-base tests. Also, ensure the query variable selector is set to pass the user's query (like sys.query) into the retrieval node—misconfiguration here can cause silent empty results (source).

• If you see no errors but still get empty results, check backend logs for exceptions or warnings during workflow execution. Errors in any retrieval thread (for example, due to incompatible settings) can cause the whole retrieval step to fail (source).

• Performance/resource issues can also cause failures when multiple knowledge bases are configured. If you notice slowdowns or timeouts, try tuning environment variables like RETRIEVAL_SERVICE_EXECUTORS (controls parallelism), database pool settings, or reduce the per-knowledge-base top_k value (source, source).

A good way to isolate the problem is to create a minimal workflow, add one knowledge base at a time, and test after each addition. This can help pinpoint which combination or setting is causing the failure.

Recent code changes have focused on retrieval efficiency and bug fixes, but there is no direct fix for this specific issue yet. Monitoring for future updates or patches is recommended (source).

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}