[Feature Request]: DocAgent: Improve Detail Retention Capabilities

[gabriel-eidelman]

Is your feature request related to a problem? Please describe.

The ability to retain details from a document is very important for various applications. DocAgent has the ability to store documents in-memory, which allows it to answer queries about the specifics of a (relatively short) document with much more accuracy. However, the final step in DocAgent's process is always to pass to a SummaryAgent, which often loses important details from the query responses. If DocAgent is being used in a groupchat for example, then it's very difficult for other agents to get any access to doc details and not just a summary. Often, it would be much more useful to simply terminate after the query step, without summarization, and then return the full set of query responses.

Describe the solution you'd like

I would propose that we add a parameter to DocAgent, perhaps called something like detail_mode which, if set to true, terminates when QueryAgent is complete and then returns as the result of the chat the full set of queries and responses.

Here is a possible solution I would propose:

(added to DocAgent init)

detail_mode: Optional[bool] = True,

near the top of DocAgent:

self.detail_mode = detail_mode

where we set the handoffs:

if self.detail_mode:
    self._query_agent.handoffs.set_after_work(target=TerminateTarget())
else:
    self._query_agent.handoffs.set_after_work(target=AgentTarget(agent=self._task_manager_agent))

Adding a custom formatting function so that the format of the queries is consistent. It could be placed in generate_inner_group_chat_reply.

def format_query_results(query_results):
    output_lines = []
    for i, item in enumerate(query_results, start=1):
        output_lines.append(f"{i}. **{item['query']}**")
        output_lines.append(f"   - {item['answer']}")
    return "\n".join(output_lines)

and finally in inner_group_chat_reply, changing the default return statement to:

if self.detail_mode:
    formatted_block = format_query_results(context_variables.get("QueryResults", []))
    return True, formatted_block
else:
    return True, chat_result.summary

This works because we're already saving the list of QueryResults in context_variables.

Additional context

No response

[marklysze]

Looks good @gabriel-eidelman, and makes sense to have.

Questions:

• Confirming detail mode works in other memory modes as well, such as with the vector DB?
• I'm not sure we would default it to True, do you think it should so?

[gabriel-eidelman]

Thanks @marklysze! Great questions

• I did test it with the vector db and it didn't cause any problems - it's not changing anything related to memory management so it shouldn't cause any issues
• Definitely right my mistake, it should default to False