I have a prompt for tool enforcement - List of tools (RAG tool specified) - when a customer asks a question - it goes through different stages of prompt - For eg Introduction, sales call, pre-sales, information collection etc - The RAG tool currently needs to be enforced only if customer asks “Can you check the knowledge verbatim” - Considering customer cant utter these words - even though the real time session update has list of tools and prompt it needs to access the tools, what’s the solution for checking the query and access the tool I mention for sure as required - since I have bunch of tools for each stage
it depends on your JSON schema you defined, it should have clear description corresponding to user query but not so specific that only specific query would trigger that tools.
Yes its clearly defined - but the probablity of model understanding the query and executing the tool is far less than the verbatim like “Can you check knowledge base for it”
Provide more detail in the description of each field in the schema.
Smaller models might not capture some implicit concepts, so make sure to use GPT-4o at least.
If it still doesn’t work, you need to manipulate the user input in another call and check against functions. For instance, classify user input into categories based on your tools’ functionalities.
If it is possible, provide more details
Yes i think there is a deep intermediate layer involved in query decomposition vs tool calling based on the query. Inherently unless “search in knowledge base” is sent out - tool calling “auto” is very fragile - Its best choice is always not to use tool calling and using its own knowledge or knowledge in system prompt