goal: Deep Research Agent POC

[Yip-Jia-Qi]

Background

menloresearch/jan-models#271

Goal

• A deep research agent that is able to orchestrate web search and report writing over multiple turns
• The Agent is used via and API written to industry standards (references OpenAI, GoogleGemini and Tongyi Deep Research)

User Stories

• User can use the same Deep Research functionality as other mainstream solution to generate a report
• User can modify details of how the Deep Research Agent works, without breaking the functionality. for example, if the user needs to use a different search engine

Requirements

There are a number of components we need for DeepResearch

1. An LLM model
2. Search Tools
3. Pipeline Logic

We should be able to configure and mix and match different variations of each of these. They may not all work, but we should build for ourselves a research tool that allows us to easily change these components.

In the future, when we make improvements to deep research we should also be able improve make standalone improvements to each of these components

Scope

• for model, you can use anything that can be found on open-router

Out of Scope

• training our own model. That is an optimisation we will make once we have users.
• Authentication
• User self-service creation of pipelines, tools and models. we will pre-define each of these components.

Open Questions

• What are the best models to use for this?
• What is the best way to orchestrate the models?
• What search tool is most feasible and scalable to use?
• Do we need to pay for search? Can we not pay for search?
• How do we measure the performance of the overall system?

[Yip-Jia-Qi]

https://github.com/menloresearch/jan-incub-agents/blob/main/attachments-agent/poc_document_retrieval/modules/main.py