DeepReason
A multi-agent, self-correcting retrieval-augmented reasoning engine that plans multi-hop evidence retrieval, then has a critic and reviser iterate on an answer until it converges, with reproducible evaluation against a judge.
DeepReason is an agentic retrieval-augmented reasoning engine orchestrated as a state machine. A planner decides, based on question complexity, whether to answer directly or split a question into sub-questions and retrieve evidence for each one across multiple hops; a three-agent loop of a generator, a critic and a reviser then reviews the draft answer, decides to accept, revise or reject each claim, and revises according to the critic's notes, stopping early once revisions converge or falling back to the earlier draft on rejection rather than emitting a worse answer.
The critic checks assertions against tools resembling MCP for paper search, concept lookup, framework documentation search and comparison, and guardrails include convergence detection based on similarity between drafts, enforced runtime limits, risk tagging, and low-confidence tagging that only affects metadata rather than the visible answer text. Evaluation uses a large-language-model judge with programmatic, weighted scoring across three configurations, baseline, draft and full, on a 30-question set stratified across papers, framework documentation and technical posts.
It is built on LangGraph with a Milvus vector store, uses a DeepSeek model for each agent role, and includes a web interface that streams the reasoning process live.