🔎 InvestigA — Fact-Checking with RAG + Ollama

An automated fact-checking system that verifies claims against a curated corpus of 19,502 Brazilian fact-checking articles using Retrieval-Augmented Generation (RAG) and local LLMs via Ollama. Developed as part of a CNPq-funded research project on combating disinformation in Brazil.

✨ Features

• 4-Step Pipeline: Question generation → Evidence retrieval → Per-question answering → Classification
• 3 Retrieval Modes: RAG (local corpus), Web (Google Search), Hybrid (RAG → web fallback)
• 3 Retrieval Methods: BM25 (keyword), Semantic (embeddings), Hybrid (combined) — switchable at runtime
• 2 Verdict Strategies: Ollama LLM classification, Label majority voting
• Multi-Run Consistency: Classification runs N times (default: 3); confidence = % agreement (not LLM self-reported)
• Per-Question Answering: Each investigative question is answered individually by the LLM using its retrieved evidence before final classification
• Runtime Model Selector: Switch between any downloaded Ollama model via webapp dropdown
• Claim History: Every analysis is saved; browse, expand Q&A details, and clear past results
• Dark Theme UI: Dark navy background with white text, all labels translated to Portuguese
• Semantic Search: Sentence-transformer embeddings with chunk & pool encoding (500-char chunks, max-sim)
• Embedding Caching: First run encodes articles (~10 min), subsequent restarts are instant
• 19,502 articles from 6 Brazilian fact-checking sources (G1, Lupa, Aos Fatos, Estadão, Boatos.org, UOL Confere)
• Web interface with real-time SSE streaming
• FastAPI backend serving both API and webapp

🚀 Quick Start

Prerequisites

• Python 3.10+ (conda env cnpq recommended)
• Ollama installed with at least one model (e.g., qwen3:14b)
• JSONL datasets in coleta_datasets/data/raw/

Setup

# 1. Install dependencies
conda activate cnpq
pip install -r requirements.txt

# 2. Start Ollama (if not running as system service)
CUDA_VISIBLE_DEVICES=0 OLLAMA_MODELS=/mnt/E-SSD/model_cache ollama serve

# 3. Pull a model (if not already downloaded)
ollama pull qwen3:14b   # or any model of your choice

# 4. Start the API server
CUDA_VISIBLE_DEVICES=0 uvicorn api:app --host 0.0.0.0 --port 8000
# ⚠️ First run encodes ~156k chunks (~10 min). Subsequent starts are instant.

Remote Access (SSH)

# On your local machine, create an SSH tunnel:
ssh -L 8000:localhost:8000 <server>

# Then open: http://localhost:8000

📁 Project Structure

InvestigA/
├── api.py                       # FastAPI server + static file serving
├── main.py                      # Pipeline orchestrator (init_rag, run_pipeline)
├── .env                         # Configuration (GPU, model, paths)
├── requirements.txt             # Python dependencies
├── pipeline/
│   ├── __init__.py
│   ├── generate_questions.py    # Ollama-powered question generation
│   ├── retrieve_evidence.py     # Multi-mode evidence retrieval dispatcher
│   ├── rag_retrieval.py         # BM25 + semantic RAG index (RAGIndex class)
│   ├── data_loader.py           # Unified JSONL corpus loader
│   ├── answer_questions.py      # Per-question answering using retrieved evidence
│   └── classification.py        # Verdict strategies (ollama_verdict, label_vote)
├── webapp/
│   ├── index.html               # Main UI with mode/retrieval/strategy/model dropdowns
│   ├── script.js                # SSE client, Q&A rendering, history panel
│   └── style.css                # Dark theme styling with source/label badges
└── data/
    ├── results.json             # Latest pipeline output (auto-generated, gitignored)
    ├── history.jsonl            # Claim analysis history (auto-generated, gitignored)
    └── embeddings_cache/        # Cached .npy embeddings (auto-generated, gitignored)

⚙️ Configuration (.env)

Variable	Default	Description
CUDA_VISIBLE_DEVICES	0	GPU to use
OLLAMA_MODEL	qwen3:14b	Default Ollama model (can be overridden via webapp dropdown)
OLLAMA_MODELS	/mnt/E-SSD/model_cache	Ollama models directory (E-SSD)
DATA_DIR	/mnt/.../data/raw	Path to JSONL dataset directories
SEMANTIC_MODEL	paraphrase-multilingual-mpnet-base-v2	Sentence-transformer model for semantic search
ENCODING_STRATEGY	chunk_pool	Encoding strategy: chunk_pool, title_label, truncate
EMBEDDINGS_CACHE_DIR	./data/embeddings_cache	Directory for cached embeddings (.npy)
MODEL_CACHE_DIR	/mnt/E-SSD/mussi/model_cache	HuggingFace/Torch model download directory
CONSISTENCY_RUNS	3	Number of LLM classification runs for confidence calculation
GOOGLE_API_KEY	—	For web/hybrid retrieval modes
GOOGLE_CSE_ID	—	For web/hybrid retrieval modes

📖 Documentation

See DOCUMENTATION.md for comprehensive technical documentation.

🔗 Data Sources

Source	Articles	File
G1 Fato ou Fake	1,907	g1_cleaned.jsonl
Lupa	4,141	lupa_cleaned.jsonl
Aos Fatos	3,537	aosfatos_cleaned.jsonl
Estadão Verifica	1,695	estadao_cleaned.jsonl
Boatos.org	6,556	boatos_2020_2025_cleaned.jsonl
UOL Confere	1,666	confere_cleaned.jsonl
Total	19,502

📜 License

Part of a CNPq-funded research project on combating disinformation in Brazil.