Preserve Original Input: Using Conversational Agents in Haystack

[Halpph]

I'm interested in using a conversational agent with Haystack, but I'd like to preserve the original prompttext without any interpretation or modification by the underlying tool. Currently, when I input a question into Haystack, it seems like the tool may interpret or change my words. Is there a way to configure Haystack to accept and process questions as they are, without any alterations to the input text? If not, could you provide guidance on how to achieve this? This is important for scenarios where preserving the exact phrasing of the user's query is crucial. Thanks for your help!

[vblagoje]

Hey @Halpph can you please provide an example of where this modification is occurring? I'm not aware of it.

[Halpph]

I'm using an agent with multiple tools to do RAG:
This is how I created my tool:

products_common_text = """Company Z products consists of the following products (Product A, Product B, Product C)"""

product_tool = Tool(
        name="CompanyProductSearch",
        pipeline_or_node=product_tool_pipeline,
        description=f"""
            Useful for finding one of the following {products_common_text} Company Z product offerings mentioned in the document.""",
        #{products_common_text}""",
        output_variable= "answers",  
    )

And this is the part of the agent prompt where I specify how to use it:

Employ the "CompanyProductSearch" tool to find one of the following <inc>{products_common_text}</inc> in the document, makes sure to include the text inside <inc></inc> as part of "Tool Input" verbatim.

This is the log of my agent run:

Tool: CompanyProductSearch
Tool Input: Which Company Z product is mentioned in the document 'Company Z Consultancy Framework Agreement Work Order signed.pdf'?

As you can see the Tool Input does not contain the names of the product I'd like to search in the original document, and I don't know how to make sure they appear in there.

[vblagoje]

Aha, not 100% sure that I understand it but I think you might resolve this by having a custom prompt_parameters_resolver given to your Agent instance during init. See how react_parameter_resolver is implemented in haystack/agents/utils.py and inject your custom keys/values where keys are template parameters. You'll likely need to repeat all the key/value pairs of react_parameter_resolver and then add your custom key/values in addition to those. LMK what happened and we'll take it from there