Hi @Logan M , how I can control how

JJeff

Hi , how I can control how many questions are generated for subquery engine and for QueryEnginetTool metadata, name is used in this process in any way or just description is takeing in consideration ?

Plain Text

# setup base query engine as tool
query_engine_tools = [
    QueryEngineTool(
        query_engine=vector_query_engine,
        metadata=ToolMetadata(
            name="pg_essay",
            description="Paul Graham essay on What I Worked On",
        ),
    ),
]

18 comments

LLogan M

It's using name + description of each query engine, plus the user query to generate sub-questions

This is actually the prompt template

Attachment

LLogan M

I don't see a way to limit questions besides editing the tempate, which is used in the question generator module

JJeff

can I use custom prompt for this ?

LLogan M

You should be able to do something like

Plain Text

from llama_index.prompts import PromptTemplate

prompts_dict = query_engine.get_prompts()
print(list(prompts_dict.keys()))

prompt_tmpl_str = ...
prompt_tmpl = PromptTemplate(prompt_tmpl_str)

query_engine.update_prompts(
    {"...": prompt_tmpl }
)

LLogan M

Source: https://docs.llamaindex.ai/en/stable/module_guides/models/prompts/usage_pattern.html#accessing-prompts

JJeff

many thanks I will try

JJeff

how each subquestion is processed ? there is a separate call to query_engine, If I use vector_query_engine and have configured to use 4 chunks, for each subquestion will retrieve different chunks and send to LLM?

LLogan M

Each sub question + response and the original user query is sent to the LLM to produce a final response

LLogan M

Each response+sub question is turned into a "node"

Attachment

LLogan M

Then all those nodes + original query is sent to the response synthesizer

Attachment

JJeff

ok, what I don't understand is if for user question I retrieve 4 chunks from my database, this 4 chunks will be used for all sub question? or for each sub-question it take different chunks ?

LLogan M

The latter

LLogan M

Think of it this way

LLogan M

Attachment

LLogan M

🧑‍🎨

LLogan M

It's running entire query engines under the hood for each sub question

JJeff

super, is more clear now, Thank you

JJeff

@Logan M I'm having some issues configuring this. I'm getting this type of messages "Unfortunately I do not have enough context in the provided tools and user question to generate relevant sub-questions.

I do not have enough information to break this down into sub-questions.

The provided tools do not seem directly related to answering the user's question about whether <my question>. To provide helpful sub-questions, I would need more details on the specific tools available and how they could assist ..... Please provide additional context about the tools and their relation to address the user's question." Are this related with information provided in description here query_engine=vector_query_engine,
metadata=ToolMetadata(
name="pg_essay",
description="Paul Graham essay on What I Worked On",
) ?

Add a reply

Find answers from the community

Hi @Logan M , how I can control how