Introduction

An overview of the Exquisitor Live Services Engine and its core capabilities.

What is the Live Services Engine?

The Live Services Engine (LSE) is the Python/FastAPI backend for Exquisitor, a multimedia search and exploration system. It exposes a REST API that powers three distinct search strategies over large image collections:

  • CLIP search — text-to-image retrieval using the ViT-SO400M-14-SigLIP-384 CLIP model
  • Relevance feedback — iterative search refinement using a linear SVM trained on user-selected positive and negative examples
  • Faceted filtering — metadata-only queries using a composable filter expression tree

The engine is designed to serve multiple independent collections simultaneously, each with its own vector index and SQLite metadata database.

Key features

  • Three search modes: CLIP text search, SVM relevance feedback, and filter-only faceted search work independently or in combination.
  • Multi-collection support: Each collection has its own configuration, index, and database. All are loaded at startup and served from a single process.
  • Flexible vector indices: Supports both brute-force Zarr arrays and approximate nearest-neighbour FAISS indices (IVF or HNSW).
  • Composable filters: A tree-based filter expression system allows arbitrary AND/OR/NOT combinations of categorical and numerical constraints.
  • Audit logging: All search requests, item views, and client events are written to a structured MessagePack log.
  • Hardware-aware: Automatically selects CUDA, Apple MPS, or CPU for model inference. The CLIP text model is cached to disk after first download.

Architecture at a glance

Client
  │
  ▼
FastAPI (main.py)
  │
  ├── /exq/search/   →  SearchService  →  Strategy (CLIP / RF / Faceted)
  ├── /exq/item/     →  ItemService
  ├── /exq/          →  Admin / Init / Filters
  └── /health        →  Health check
                            │
              ┌─────────────┴─────────────┐
              ▼                           ▼
      IndexRepository             DatabaseRepository
      (FAISS / Zarr)              (SQLite per collection)

What's next?