FDA Molecule Intelligence Agent

A local-first AI pipeline that extracts verified medical facts from FDA drug labels (PDFs) and indexes them into a searchable semantic knowledge base. It uses small, efficient LLMs (Gemma-2 2B) running on CPU to parse complex unstructured data into structured, cited intelligence.

🚀 Key Features

• Local Privacy: Runs entirely offline using ollama. No data leaves your machine.
• Dual-Engine Storage:
  • Relational (SQLite): Stores high-fidelity facts, audit trails, and performance metrics.
  • Semantic (ChromaDB): Enables conceptual search (e.g., "Find drugs with renal risks") using vector embeddings.
• Hallucination Guardrails: A deterministic QuoteVerifier ensures every extracted fact is backed by an exact substring match in the source text.
• Audit Trail: Every LLM thought, prompt, and latency metric is logged for full reproducibility.

🏗️ Architecture

The system follows a two-stage RAG (Retrieval-Augmented Generation) pipeline:

1. Extraction Layer: PDFs ⮕ LLM ⮕ SQLite (Facts + Quotes)
2. Semantic Layer: SQLite ⮕ Embedding Model (all-MiniLM-L6-v2) ⮕ ChromaDB

🛠️ Setup

1. Prerequisites

   • Python 3.10+
   • Ollama installed and running.
   • Poetry (Python dependency manager).

2. Installation

   # Install dependencies via Poetry
   poetry install
   
   # Pull the LLM
   ollama pull gemma2:2b

3. Usage Flow

   Step 1: Extract Facts

   Process your PDFs to populate the SQLite audit store.

   poetry run python -m src.main data/raw_pdfs/keytruda.pdf

   Step 2: Build Knowledge Base (New)

   Convert extracted facts into semantic vectors for searching.

   poetry run python -m src.scripts.build_knowledge_base

   Step 3: Semantic Query

   Ask the agent conceptual questions.

   poetry run python -m src.scripts.query_agent

🛠️ Tech Stack

• Orchestration: Python 3.10 + Poetry
• LLM: Gemma-2 2B (via Ollama)
• Databases: SQLite (Audit/Facts), ChromaDB (Vector Store)
• Embeddings: Sentence-Transformers (all-MiniLM-L6-v2)
• UI: Rich (Terminal formatting)