Issues of Retrieval Accuracy and Document Support

[zephyershen]

Recently, when using v0.18.0-slim I found that the document parsing for DOC-format files in my database was successful. However, when I ask questions about the content of these documents, the answers cannot find the relevant information. Additionally, the accuracy is not very high when asking questions about the content of other format documents sometimes.

I will provide the configuration of my database and chat settings, and request help to solve the problem!

[zephyershen]

The content I asked about is exactly copied from the DOC document, but the response says it cannot be found.

[zephyershen]

Always showing this problem, unable to enter @ZhenhangTung

[ZhenhangTung]

@zephyershen It's WeChat's risk control policy. Maybe you added too much WeChat groups recently? Wait for hours and then retry.

[ZhenhangTung]

@zephyershen What's your WeChat ID or nick name?

[zephyershen]

From yesterday until today, it has been showing that the operation is too frequent，Here is my WeChat ID: MrShenzf @ZhenhangTung

[zephyershen]

Here is my WeChat ID: MrShenzf @ZhenhangTung

[zephyershen]

I need the knowledge base to return all the information, but it is always incomplete. How can I solve this?

[zephyershen]

Why does the knowledge base contain this information, but when I ask about it, it says it can’t be found? For example, when I ask what the content of the function PPP_CHAP_input is, the answer is that it can’t be found, even though the information actually exists. @dosu

[dosubot[bot]]

The issue where the knowledge base contains information (such as the function PPP_CHAP_input) but cannot retrieve it when queried is most likely related to how RAGFlow processes, indexes, and retrieves content. Here are the key factors and solutions based on your configuration and RAGFlow's retrieval logic:

1. Retrieval Process and Query Matching

RAGFlow uses a combination of weighted keyword similarity and vector cosine similarity to match your query against document chunks in the knowledge base. If the similarity score between your query and the indexed chunks is below the set threshold, those chunks will not be returned—even if the information exists in the database. This can happen if:

• The function name or technical terms are not well-represented in the embeddings.
• The segmentation or chunking of the document splits relevant information across boundaries, making it harder to match your query to a single chunk.
• The similarity threshold is set too high, filtering out relevant but not perfectly matched chunks.

You can review and adjust these parameters:

• Similarity threshold: Lowering this value can help include more potentially relevant chunks in the results.
• Keyword similarity weight: Adjust the balance between keyword and vector similarity to better match technical queries. The default is 0.7 for keywords and 0.3 for vectors, but you can experiment with these values for your data.
• Top N: Increasing this value will allow more chunks to be considered by the LLM, which can help with coverage but may reduce precision. The default is 8. Reference

2. Segmentation and Chunk Size

If the function definition or description is split across multiple chunks due to the segmentation method or block size, the retrieval may miss the full context. Try:

• Reducing the text block size to create smaller, more focused chunks.
• Experimenting with different segmentation methods to see which best preserves function boundaries.

3. RAPTOR Strategy

The RAPTOR strategy clusters and summarizes document chunks recursively, which can help with long or complex documents. However, if not tuned properly, it may also abstract away details needed for technical queries. You can adjust:

• max_token: Controls the size of each chunk/summary (up to 2048).
• threshold: Controls the similarity threshold for clustering (0–1).
• max_cluster: Controls the number of clusters (up to 1024).
• random_seed: For reproducibility.
  Reference

4. Performance and Feature Tuning

• Disabling the reranker and auto-keyword/auto-question features can improve performance and sometimes retrieval quality, especially on CPU.
• Using a GPU can speed up embedding and retrieval.
• For plain text documents, using the "Naive" parser instead of "DeepDoc" may yield better results if your documents are not highly structured.
  Reference

5. Code Reference

The retrieval logic is implemented in the KGSearch class, which rewrites queries into keywords/entities and retrieves relevant entities and relations from the knowledge base. If your query is not being rewritten effectively (for example, if "PPP_CHAP_input" is not recognized as an entity or keyword), it may not match the indexed content. Reference

---

Summary of Actions to Try:

• Lower the similarity threshold and increase "Top N" in your retrieval settings.
• Adjust the keyword/vector similarity weights.
• Reduce text block size or try different segmentation methods.
• Tune RAPTOR parameters or disable RAPTOR to compare results.
• Ensure the function name appears as a distinct chunk in your indexed data.
• Consider disabling reranker and auto-keyword features for testing.

If after these adjustments the issue persists, it may be necessary to review how your DOC files are parsed and indexed, or to check if the function names are being tokenized or embedded in a way that matches your queries. If you can share a sample document (with sensitive info removed), it would help with more targeted troubleshooting.

To continue talking to Dosu, mention @dosu-bot.

---

^{Help Dosu learn! Give it feedback: Great Response | Irrelevant Answer | Incorrect Sources | Too Verbose | Hallucination | Bug Report | Other}