Autonomous research agent: given a query, it plans sub-queries across ArXiv, Tavily, Wikipedia, and SerpAPI, synthesizes a cited report, detects conflicts, self-critiques, and refines until satisfied. Results stream in real time to a chat UI. Config: config/config.yml for defaults, .env for secrets and URLs.

• Multi-source search — ArXiv, Tavily, Wikipedia, SerpAPI with dynamic source selection
• Query expansion — Planner produces 3–5 focused sub-queries per iteration
• Conflict detection — Synthesizer highlights contradictions across sources
• Self-critique loop — Critic scores quality; loops back to Planner if score < 0.7
• Long-term memory — ChromaDB for past reports and source credibility
• Real-time streaming — SSE for steps, sources, and report text
• Checkpointing — Postgres for recoverable LangGraph state
• Distributed mode — Optional Celery + Redis for high concurrency

User submits a query → Frontend starts a task and opens an SSE stream. Planner generates sub-queries → Workers search in parallel → Synthesizer builds a cited report and detects conflicts → Critic scores the report; if below threshold, loop back to Planner with feedback. When done, the report is stored in ChromaDB and streamed to the client.

Samwel Ngusa