Qwen3 does not seem to work well with file_search tool, doesn't recognize the user's query

[enricbaixauli]

What happened?

Hello, I am experimenting with librechat for RAG and I tried Qwen3 for my personalized agent with file-search. The problem is that when I ask the agent something related with the documents that it ingested it seems to not find my query and it only answers about all the chunks from the file_search tool:

However, when the agent does not use the tool, it works fine and actually knows what I am requesting:

I have this instruction prompt for my agent: "You are an assistant at #. Use the file_search tool to search the documents you have and prepare a response based on them in Spanish. If you do this in English, you will translate it into Spanish before answering. You will always respond only to what is requested; you will not digress. You will not refer to the documents in your response; you will respond only with the information you have gathered from the files. You will only answer in spanish."

Thank you in advance, I am loving this project so far.

Version Information

I am using LibreChat v0.7.8 and qwen3:30b-a3b

Steps to Reproduce

Download qwen3 from ollama
Create an agent with ollama and select the model, give him instructions to use search_file
Feed the agent with files
Ask the agent a question about the files

What browsers are you seeing the problem on?

No response

Relevant log output

.

Screenshots

No response

Code of Conduct

• I agree to follow this project's Code of Conduct

[danny-avila]

It's an LLM issue. The output of what it's getting can also largely sway how it responds.

However, qwen3:30b-a3b-q8_0 did a decent job using file search for me:

[enricbaixauli]

Hi,
It's weird because when I ask the model and I give it only one document it works well:

But when I use all the documents it still doesn't recognize the query:

I've tried llama4 as well and the same thing happens:

Thanks for answering.

[enricbaixauli]

Pd: the chunks that the file_search retrieves are of good quality

[danny-avila]

@enricbaixauli Some local models have quite limited context windows, double check if that's the case.

[tauven]

@enricbaixauli
If you use ollama check that you configured ollama to allow more than 4000 context. (/set parameter num_ctx 32000)
Or limit the results the rag api returns add RAG_API_TOP_K=3 to your LibreChat .env file.