README.md

metadata

title: deepfake-fastapi
emoji: 🚀
colorFrom: blue
colorTo: green
sdk: docker

Model FastAPI

FastAPI service that provides two capabilities:

• Deepfake image detection using a SigLIP image-classification backbone with a LoRA adapter.
• A news-aware chatbot that selectively performs web search via Tavily and responds with verified sources when evidence is found.

Project Layout

• main app entry: main.py
• Deepfake detector: app/detector.py, app/core/detector/, app/services/detector/, model/output/siglip-lora-optimized/
• Chatbot: app/chatbot.py, app/core/chatbot/, app/services/chatbot/, app/schemas/chat.py
• Database migration (chat history table): supabase/migrations/

Requirements

• Python 3.13+
• uv package manager (https://docs.astral.sh/uv/)
• Access to the SigLIP base model and the LoRA adapter stored at model/output/siglip-lora-optimized/

Environment Variables

• DATABASE_URL (or SUPABASE_DATABASE_URL): PostgreSQL connection string for chat history (if save_to_db=true).
• GROQ_API_KEY: required by the Groq LLM used for responses.
• OPENROUTER_API_KEY: required by the OpenRouter model used for classification/query rewriting.
• TAVILY_API_KEY: required for Tavily search.
• Optional: set CUDA-visible devices as needed for GPU inference.

Setup (local)

uv sync
uv run main.py  # starts FastAPI on 0.0.0.0:7860

The detector will load the SigLIP base model and apply the LoRA adapter from model/output/siglip-lora-optimized/.

API

• POST /detect
  • Form-data file field: file (image). Returns predicted_class (index), predicted_label (from id2label), prediction (real/fake thresholded at P(real) >= 0.90), confidence, and class probabilities.
• POST /chat
  • JSON: {"query": "...", "session_id": "optional", "save_to_db": true|false}
  • Auto-classifies need for search, optionally queries Tavily, then responds. Returns response.content, session_id, used_search, and search_reason.
• DELETE /chat/{session_id}
  • Clears chat history (both in-memory guest sessions and DB rows).

Quick cURL examples

# Deepfake detection
curl -X POST "http://localhost:7860/detect" \
  -F "file=@path/to/image.jpg"

# Chat (without forcing search decision)
curl -X POST "http://localhost:7860/chat" \
  -H "Content-Type: application/json" \
  -d '{"query": "Is the latest SpaceX launch successful?", "save_to_db": false}'

Docker

A Dockerfile is provided using the uv base image.

docker build -t model-fast-api .
docker run -p 7860:7860 --env-file .env model-fast-api

The image installs dependencies with uv sync and runs uv run main.py.

Notes

• The news verification prompt only returns source links when the claim is supported by the retrieved results; otherwise it replies with UNDETERMINED and no links.
• If chat history persistence is disabled (save_to_db=false), sessions are stored in-memory.