IntraIntel.ai - Multi-LLM Agent Coding Challenge

What is IntraIntel.ai - Multi-LLM Agent Coding Challenge?

It is a Python-based multi-agent system that answers medical questions by orchestrating multiple free LLMs from the Hugging Face Inference API. It uses a custom HTTP-based Model Context Protocol (MCP) to interact with separate tool servers for web search (DuckDuckGo) and PubMed search, and implements a Retrieval-Augmented Generation (RAG) pipeline with optional query refinement and snippet summarization.

How to use IntraIntel.ai - Multi-LLM Agent Coding Challenge?

Clone the project, create a Python virtual environment, install dependencies (FastAPI, uvicorn, httpx, duckduckgo-search, biopython, python-dotenv), and set HF_API_TOKEN and NCBI_EMAIL in a .env file. Accept the license terms for mistralai/Mistral-7B-Instruct-v0.3 on Hugging Face. Start the Web Search MCP server on port 8001 and the PubMed MCP server on port 8002 in separate terminals. Then run python agent/main_agent.py to process a predefined set of medical questions.

Key features of IntraIntel.ai - Multi-LLM Agent Coding Challenge

• Multi-step RAG pipeline using up to three distinct LLMs.
• Optional query refinement and snippet summarization steps.
• Separate synthesized answers from web search and PubMed with source links.
• Asynchronous HTTP communication via FastAPI and Uvicorn.
• API-key-free web search via DuckDuckGo and free Hugging Face models.

Use cases of IntraIntel.ai - Multi-LLM Agent Coding Challenge

• Answering medical questions by combining general web and PubMed literature.
• Comparing synthesized answers from two independent information sources.
• Automating evidence-based research for clinical or educational queries.
• Demonstrating an agentic RAG workflow with multiple free LLMs.

FAQ from IntraIntel.ai - Multi-LLM Agent Coding Challenge

Which LLMs are used?

The system uses mistralai/Mistral-7B-Instruct-v0.3 for query refinement and answer synthesis, and sshleifer/distilbart-cnn-6-6 for snippet summarization, both accessed via the free Hugging Face Inference API.

What are the main dependencies?

FastAPI, uvicorn, httpx, duckduckgo-search, biopython, and python-dotenv.

How do the MCP servers communicate?

They use a simple HTTP-based JSON protocol: a POST request to /execute with {"query": "..."} returns a JSON response containing source, status, and results (title, snippet, URL or PubMed ID).

What authentication is required?

A Hugging Face API token (HF_API_TOKEN) with read permissions, and an email address (NCBI_EMAIL) for polite access to NCBI’s Entrez system.

Are there performance considerations?

The first call to a Hugging Face model may take 20–90 seconds as the model loads; subsequent calls are faster. The script includes pauses to respect API rate limits.