vector-search-hf

Minimal TypeScript vector search service — embeds text with HuggingFace models, stores in ChromaDB, and serves semantic search via an Elysia HTTP API.

A small reference implementation for building a vector search backend with:

• HuggingFace Inference API (or local models) for text embeddings
• ChromaDB as the vector store
• Elysia (Bun's fast web framework) for the HTTP API

Stack

• Runtime: Bun
• Language: TypeScript
• Server: Elysia
• Vector store: ChromaDB
• Embeddings: HuggingFace models

Getting started

bun install
docker-compose up -d   # start Chroma
bun run dev

Environment

HF_API_KEY=         # HuggingFace API token
CHROMA_URL=http://localhost:8000

Project layout

src/
├── embedding/   # HF embedding client
├── types/       # shared types
└── index.ts     # Elysia server + routes
docker-compose.yml   # Chroma

Why

Built as a reference for hooking LLMs into domain data: pick an embedding model, ingest your documents, and query by similarity. Minimal surface area so it's easy to adapt to other vector stores (pgvector, Qdrant, etc.).