Dify Knowledge Base Test issues when building bot assistants

[mianmianbing-design]

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

I used the Dify knowledge base to design a process. My goal is to create a robot assistant that can answer questions based on the knowledge base. I designed a workflow where the user inputs a question, and the LLM first interprets it using the chat model, then searches through the knowledge base documents, and finally provides an answer to the user via the same chat model. Additionally, I configured the knowledge base recall settings to use the rerank model BGE-Reranker-v2-m3 along with the embedding model, both of which are deployed locally. I also set the document segmentation mode to parent-child.

However, I encountered some issues during testing:

1. Even though the relevant documents have been uploaded to the knowledge base, they don’t seem to match at all during the recall test. Could you help me understand why this might be happening?
2. How can the robot better identify specific keywords in the knowledge base when users ask questions, so that it can find the most relevant information?
3. What improvements would you suggest for my current setup?

Thank you very much for taking the time to read through my questions. As someone who is just starting out, I truly appreciate any guidance or suggestions you may have.

[dosubot[bot]]

Hi @mianmianbing-design! I'm Dosu and I’m helping the dify team.

Here are some common reasons why your uploaded documents might not be matching during recall in Dify:

• Parent-child segmentation requires careful configuration. Make sure your chunking rules (like max_tokens, chunk_overlap, and separators) are set appropriately for both parent and child nodes, and that you’re using the High-Quality Indexing method. Misconfiguration here is a frequent cause of recall failures (source).
• If one document is much larger or more relevant, it can dominate recall results, making it hard for other documents to appear. Try segmenting large documents, increasing the top_k value, or using metadata filtering to improve recall diversity (source).
• Content mismatches can result from Dify’s chunking and cleaning steps, which may alter text format. If you notice missing or changed content, switch from automatic to custom chunking and cleaning rules to better preserve your original text (source).
• Minor changes in input (like whitespace) can affect embeddings and retrieval results, especially with the BGE models (source).
• Double-check that your embedding and rerank models are enabled and correctly set as defaults. Sometimes Dify selects a disabled model by default, which can silently break recall (source).

To improve keyword identification and retrieval accuracy, enable keyword search in your retrieval model settings (recent updates added explicit support for this), and consider adding metadata to your documents for more precise filtering (source). Adjust chunk sizes and overlaps to ensure important keywords aren’t split across chunks.

For best results: review your segmentation and retrieval settings, use metadata filtering, and validate your model configurations. If you’re still not seeing matches, try uploading a small test document and querying for an exact phrase to confirm the pipeline is working end-to-end. If you need more specific help, sharing your chunking settings and a sample document would help narrow things down.

To reply, just mention @dosu.

---

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