[BUG] Retriever Tool called with parameter 'query' instead of 'input' - Received tool input did not match expected schema (Reproducible!) about flowise HOT 4 CLOSED

Some other community member bumped into this issue as well. In short, gpt-4o doesnt quite follow the function calling input.

Here's the solution:
"Tool Agent + gpt4o, where it would sometimes use "query" as input key, where the system expects "input" as input key. The solution for me was adding: Please ensure that when calling this tool, the input key MUST always be "input". to each tool description and adding When calling any tool, the input key MUST always be "input". to the priming instructions on Tool Agent. Hope this helps for you as well."

from flowise.

Hi @AlehandroL, I don't really understand the problem you're facing and I can't replicate it.

• In LangSmith our code 9 months ago we already using "inputs".
• In Retriever Tool our code 4 months ago we already using "inputs".

This is my chatflow:

• issues_2552 Chatflow.json

These are my sample data:

• Bill Gates.txt
• Elon_Musk.txt

Results:

Could you share your custom tool to me? so I can debug further.

from flowise.

Some other community member bumped into this issue as well. In short, gpt-4o doesnt quite follow the function calling input.

This is the problem I tried (poorly) to describe.

The solution on the side of descriptions seems to work in many cases, but I think you could catch this error and try to look for the “query” parameter in the output of the LLM. I know it’s not ideal, but considering that GPT-4o today is (arguably) the most popular model, it’s worth considering a solution for this issue.

@HenryHengZJ

from flowise.

added a potential fix - #2679, closing for now

from flowise.