Open Source Implementation Of Deep Research Agent

In this project, the following Agentic topology is implemented:

1) A user provides a query or topic to be researched.

2) A large language model (LLM) generates an outline for the final report, limited to a specified number of paragraphs.

3) Each paragraph description is individually processed through a research phase to gather comprehensive information for report construction. (Details of this research phase are described in the next section.)

4) The gathered information is passed through a summarization step to generate the final report, including a conclusion.

5) The report is delivered to the user in Markdown format.

Each research step follows this process:

1) Based on the paragraph outline, a LLM is prompted to generate web search queries designed to retrieve the most relevant supporting information.

2) The LLM outputs the search query along with its reasoning.

3) The query is executed against a web search engine, and top relevant results are retrieved.

4) The results are evaluated through a reflection step, where the LLM identifies any missed nuances and suggests improved queries for deeper enrichment.

5) This process is repeated n times to build a robust and comprehensive information set.

Detailed walkthrough: https://www.newsletter.swirlai.com/p/building-deep-research-...