Person Search

A personal project for text/image similarity search on person datasets. It embeds images/text with SigLIP and indexes them in Qdrant for fast retrieval. Includes a Vite UI and a minimal Streamlit UI that call the same Fast API.

Dataset used: https://wanfb.github.io/dataset.html (VC-Clothes)

Video for the system

system.test.mp4

Before Run (new dataset)

1. Add a dataset class in backend/dataset.py (must return DatasetInfo with samples and metadata).
2. Update get_dataset() in backend/index_dataset.py to use your class and dataset path.
3. If you change model/collection/dim, update these in backend/index_dataset.py:
   • collection_name
   • model_name, dataset
   • embedding_dim (must match the model output)
4. Keep .env in sync with backend/index_dataset.py (model + collection).

Config

Root .env keys: DATASET_ROOT, MODEL_NAME, MODEL_DATASET, EMBED_BATCH_SIZE, QDRANT_COLLECTION, QDRANT_HOST, QDRANT_PORT, API_BASE_URL

Vite .env: VITE_API_BASE_URL

Index

1. Start Qdrant: docker compose -f backend/quadrant/dev_env/docker-compose.yaml up -d
2. Index: python backend/index_dataset.py

Run (conda)

1. Env: conda create -n person-search python=3.11 conda activate person-search
2. Configure .env (repo root) and frontend/vite/.env.
3. Backend: pip install -r requirements.txt uvicorn backend.fast:app --reload --port 8000
4. Vite UI: cd frontend/vite npm install npm run dev
5. Streamlit UI: cd frontend/streamlit pip install -r requirements.txt streamlit run app.py