infiniflow / ragflow

RAGFlow is an open-source RAG (Retrieval-Augmented Generation) engine based on deep document understanding.
https://ragflow.io
Apache License 2.0
17.19k stars 1.75k forks source link

[Question]: Decreased quality of answers and confusing references after multiple Q&A sessions in a single session window #1989

Closed Zhangli20 closed 6 days ago

Zhangli20 commented 4 weeks ago

Describe your problem

Locally deployed docker version. When using it, I found that when doing multiple rounds of Q&A in the same session window, the answers become not based on the content in the knowledge base, but on the original knowledge of the qwen2:7b, and the citations are messed up (citing snippets that were cited in other historical sessions and are not relevant to the current question). However, a new session is opened to re-ask the question (same assistant and same knowledge base) and the answer is again faithful to the knowledge base and the citation is correct.

Relevant snippets from the knowledge base 原文

Answers during multiple rounds of questions and answers 多轮问答中

First Q&A in a new session window 新会话中

Such questions occur frequently, and answers in a new session are significantly better than multiple Q&A sessions in a single session window.

KevinHuSh commented 3 weeks ago

All the chat history will be transfered to LLM. I guess LLM thought the first answer was not good enough so came the same quesion again, and it's also confused how to answer that.