A project built for IEM HackOsis 2.0 🚀 under problem statement HOGN02.

This system detects and flags hallucinations in AI-generated answers in real time, by verifying them against a trusted knowledge base (corpus.json) and optionally Wikipedia (fallback). It combines retrieval, confidence scoring, rationales, and a counter-evidence engine with a professional Streamlit dashboard for interactive analysis.

1. Start backend: cd backend && pip install -r requirements.txt && uvicorn app:app --host 127.0.0.1 --port 8000
2. Start frontend: cd frontend && pip install -r requirements.txt && python -m streamlit run streamlit_app.py
3. Try:
   • Q: "Who founded SpaceX?"
   • A: "Jeff Bezos founded SpaceX." Expect: Unverifiable/Suspected, rationale + evidence + counter-evidence.

HALLUCINATION-DETECTOR/
│
├── backend/                  # FastAPI backend
│   ├── app.py                 # Main FastAPI app (routes: /, /health, /verify)
│   ├── counterevidence.py     # Counter-evidence generation logic
│   ├── retrieval.py           # Retrieval engine (local corpus + wiki fallback)
│   ├── scoring.py             # Confidence scoring + rationale generation
│   ├── test_backend.py        # Backend test script
│   ├── utils.py               # Utility helpers
│   └── requirements.txt       # Python dependencies for backend
│
├── data/                     # Knowledge base and caches
│   ├── corpus.json            # Main trusted corpus (facts, tech, history, markets)
│   └── wiki_cache.json        # Cached wiki lookups for efficiency
│
├── docs/                     # Documentation
│   ├── PRD.md                 # Product Requirement Document
│   └── README.md              # Repo guide (this file)
│
├── frontend/                 # Streamlit frontend
│   ├── streamlit_app.py       # Streamlit dashboard UI
│   └── components/            # (Optional) Custom components
│
└── .gitignore                # Ignored files & folders

• Retrieval & Verification — Fetches supporting evidence from trusted corpus or Wikipedia.

• Confidence Scoring — Computes semantic similarity and keyword coverage.

• Short Rationale — Explains why an answer is marked as verified/suspected/unverifiable.

• Counter-Evidence Engine — Provides contradictory evidence (bonus feature).

• Streamlit Dashboard — Polished UI with:

  • Confidence & coverage bar charts 📊
  • Trend lines across multiple runs 📈
  • Evidence & counter-evidence panels 🔍
  • Session history tracking 🗂️

• Health & Root Endpoints — Quick visibility of backend status, corpus size, and model info.

1. User Input: Question + AI-generated answer.

2. Backend (FastAPI):

   • Retrieves evidence from corpus.json.
   • Falls back to Wikipedia if missing.
   • Scores confidence using SentenceTransformers.
   • Generates rationale and (optionally) counter-evidence.

3. Frontend (Streamlit):

   • Displays verdict (✅ Verified / ⚠️ Suspected / ❌ Unverifiable).
   • Visualizes confidence vs. coverage.
   • Provides trend tracking across session runs.

cd backend
python -m venv venv
source venv/bin/activate   # On Windows: venv\Scripts\activate
pip install -r requirements.txt
uvicorn app:app --reload

cd frontend
python -m venv venv
source venv/bin/activate   # On Windows: venv\Scripts\activate
pip install -r ../backend/requirements.txt
streamlit run streamlit_app.py

Unlike many heavy hallucination detectors, this project is lightweight, real-time, and explainable, with a counter-evidence module that not only checks answers but also shows why they are wrong.

This makes it:

• ⚡ Fast enough for hackathon-scale real-time QA.
• 🔍 Transparent for judges & mentors.
• 🛡️ More robust than simple similarity checkers.

Developed as part of IEM HackOsis 2.0 under Problem Statement HOGN02.

Team Goal: Build a system that balances speed, accuracy, and explainability while remaining scalable and practical.

✨ Built with: FastAPI · SentenceTransformers · Streamlit · Wikipedia API