AI Research Agent: Multi-Expert Grounding System

💼 The Business Problem: The Quest for Truth in the Age of Information Overload

Experts (researchers, decision-makers) face a triple challenge: technical data fragmentation (ArXiv vs PapersWithCode), the trust paradox (LLM hallucinations), and limited access to structured public data.

🏗️ Backend Engineering & Architecture

The heart of the system is a complex state graph developed with LangGraph, designed for performance and precision.

• •Scientific Parallelism (Fan-out/Fan-in): For scientific queries, the backend simultaneously triggers ArXiv and PapersWithCode nodes, merging theory and code into a single asynchronous stream.
• •Data.gouv Integration via MCP: Using the Model Context Protocol to dynamically query French government APIs (budgets, official statistics) with on-the-fly argument cleaning.
• •Hybrid RAG & Asynchronous Vision: Intelligent scraping via Crawl4AI, ephemeral vector storage in FAISS, and visual document analysis for perfect contextual grounding.

✨ Production-Oriented Features

• •Granular Streaming (SSE): The interface displays the agent's thinking steps in real-time (Reasoning Panel) and status events (source selection, research start).
• •Robustness & Observability: Recursion guard to prevent infinite loops and Ragas evaluation pipeline to measure response faithfulness.

🛠️ Technical Stack

• •Framework: FastAPI (Python 3.12).
• •Intelligence: LangGraph, LangChain, OpenRouter (Gemini 1.5 Flash).
• •Data: SearXNG, MCP Servers, PostgreSQL, FAISS.