Retina

Retina is a local visual search app built on CLIP embeddings, FAISS, and a small trainable query adapter.

It supports text-to-image, image-to-image, and profile-based search over a frozen CLIP backbone and a FAISS index. The repo includes training, evaluation, and a browser UI so the whole flow can be tried locally.

Architecture

flowchart LR
    A[Text / image / profile query] --> B[CLIP encoder + query adapter]
    B --> C[(FAISS index)]
    C --> D[Ranked image results]
    D --> E[React browser UI]
    D --> F[Evaluation and failure reports]

Loading

What it includes

• Query adapter trained on Flickr8k captions
• FAISS-based retrieval
• FastAPI backend and React frontend
• Gradio demo
• Evaluation metrics and failure analysis

Quick start

make install
make prepare-data
make embeddings
make train-query-adapter
make index
make eval
make api
make frontend

Summary

Retina turns a local image catalog into a searchable product with measured retrieval quality and low-latency ranked results.

Portfolio Proof

• Architecture and evaluation: docs/PORTFOLIO_PROOF.md
• UI demo and result inspector: docs/UI_DEMO.md
• Model comparison summary: docs/MODEL_COMPARISON.md
• Verified metrics: the recall and latency numbers in the proof doc
• Demo and local mode: use the make commands above
• Test commands: pytest, npm run build