HKUDS
[Question]: Poor Answering Performance
Do you need to ask a question?
- [ ] I have searched the existing question and discussions and this question is not already answered.
- [x] I believe this is a legitimate question, not just a bug or feature request.
Your Question
Hi there,
Thank you for the excellent work. I managed to setup Lightrag Server and it works fine. However I am facing an issue with the quality of the outcomes. e.g. I asked the question "List the GIAS principles, domains and standards" and used all available Query Mode, however, the best result was around 70%. The document is around 120 pages. But the answers are in two pages at max (page 3 and 4)
Any suggestions to improve the outcomes??
Additional Context
I am using LM Studio, text-embedding-bge-m3 (quantization 8). and the LLM I tried gemma-3-4b-it and qwen3-4b.
globalinternalauditstandards_2024.pdf
A 4B parameter LLM is insufficient for document indexing and querying tasks. It is recommended to use a model with at least 32B parameters and a 32K context window. Generally, larger models provide better performance.
A 4B parameter LLM is insufficient for document indexing and querying tasks. It is recommended to use a model with at least 32B parameters and a 32K context window. Generally, larger models provide better performance.
That's far away from consumer grade PC.
What about the embbeding model text-embedding-bge-m3?
The quality of the outcomes are improved after detecting below bug. I am not sure where is it coming from.
The issue is taking place with all embedding modules including bge-m3.
I used Ollama for the embedding and I think the issue is resolved., maybe expert user may verify if this issue is limited to the LM Studio and the Ollama is free from this issue.
Reranker support has been introduced in v1.4.2. With reranker enabled, the default query mode automatically switches from hybrid to mixed, significantly enhancing query performance.