Interactive search across large-scale image and video collections

Exquisitor is a research platform for interactive search over large image and video collections.

Get Started View Source Read the Docs

Technical Spec

SystemExquisitor

TypeInteractive Image and Video Search

Components2 (Frontend + Backend)

TransportREST API

LicenseAGPL-3.0-or-later

Status● Active

Overview

CLIP Text Search

Describe what you're looking for in natural language. The SigLIP model encodes your query into a 1152-dimensional embedding and retrieves the most visually similar items.

Relevance Feedback

Mark results as relevant or irrelevant. The engine trains a linear SVM on your selections and uses the learned weight vector as a new query, converging on intent across sessions.

Faceted Filtering

Compose metadata constraints with arbitrary AND/OR/NOT trees. Faceted queries compile directly to SQL.

Flexible Indexing

Choose Zarr for exact nearest-neighbour search on collections up to a few hundred thousand items, or FAISS (IVF/HNSW) for approximate search across millions with lower latency.

Research-First Design

Clean, readable Python/FastAPI code designed to be understood and modified. Each search mode is an independent strategy class: swap models, change ranking, or add new modalities.

Open Source

Released under AGPL-3.0-or-later. Use it in your research, cite it in your papers, fork it for your own experiments.

Components

All components are developed in the open. Stability labels are precise: ALPHA means the API surface will change without notice; STABLE means semantic versioning is in effect.

Repository Component Status

live-services-engine Backend / Search Engine alpha

base-user-interface Frontend / UI alpha

Simplified-M3-DB Data Model / Collection Ingest alpha

Architecture

Frontend Client

Browser-based UI for search interaction and relevance feedback

REST API

HTTP endpoints via FastAPI: search, items, admin, health

Live Services Engine

Python/FastAPI backend with CLIP, SVM, and faceted search strategies

Feature Store

Zarr or FAISS vector index + SQLite metadata database per collection

Two-component system

Exquisitor deliberately avoids microservice complexity. The entire system comprises a frontend client and a Python/FastAPI backend (the Live Services Engine) connected via a REST API. This simplicity is the point: researchers can trace every interaction from click to result.

Frontend handles all user interaction and result display
Backend manages indexing, ranking, and model inference
REST API with dedicated endpoints for search, items, and admin
FAISS or Zarr vector index chosen per collection
SQLite database per collection for fast metadata queries
Single-machine deployment, GPU optional but recommended

Quickstart

terminalbash

# Clone the Live Services Engine
$ git clone https://github.com/exq-hub/live-services-engine
$ cd live-services-engine

# Install dependencies (requires Python 3.13+ and uv)
$ uv sync

# Configure a collection
$ cp config.example.ini data/config.ini
# Edit data/config.ini to point at your index and database

# Start the server (downloads CLIP model on first run)
$ uv run python main.py

# Verify: http://localhost:8000/health
          

Getting started

Requires Python 3.13+ and uv. Point the config at your index and database, then start the server. The CLIP model is downloaded and cached on first run. A GPU is recommended but not required.