Skip to content

CLAUDE.md

Latest commit

 

History

History
55 lines (37 loc) · 1.85 KB

CLAUDE.md

File metadata and controls

55 lines (37 loc) · 1.85 KB

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

Overview

This is a TypeScript-based image search application using OpenAI's CLIP model for semantic image search. It uses the @xenova/transformers library to run a quantized version of CLIP directly in Node.js.

Architecture

The application consists of two main components:

  1. Indexing: Recursively scans directories for images, generates CLIP embeddings, and stores them in db.json
  2. Search: Takes text queries, generates text embeddings, and finds most similar images using cosine similarity

Key architectural decisions:

  • Uses singleton pattern for CLIP model management (CLIPSingleton class) to ensure models are loaded only once
  • Embeddings are persisted in a simple JSON database (db.json)
  • Supports recursive directory scanning for images
  • Implements semantic search with configurable similarity threshold (0.28) and top-N results

Development Commands

# Install dependencies
npm install

# Run TypeScript directly (for development)
npx ts-node search.ts index ./path/to/images [limit]
npx ts-node search.ts search "your search query" [top_n]

# Compile TypeScript to JavaScript
npx tsc

# Run compiled JavaScript
node dist/search.js index ./path/to/images [limit]
node dist/search.js search "your search query" [top_n]

Configuration

Key configuration constants in search.ts:

  • MODEL_NAME: 'Xenova/clip-vit-base-patch32'
  • DB_PATH: './db.json'
  • IMAGE_EXTENSIONS: ['.png', '.jpg', '.jpeg', '.webp']
  • SIMILARITY_THRESHOLD: 0.28 (minimum score to show results)

Dependencies

  • @xenova/transformers: CLIP model implementation
  • TypeScript dev dependencies: typescript, ts-node, @types/node

Important Guidance

  • Never add co-authored by Claude or any similar language to the comments or code base