README.md

---
title: Hub Search v3 Dev
emoji: 🔎
colorFrom: indigo
colorTo: blue
sdk: docker
app_port: 7860
pinned: false
private: true
short_description: DuckDB vector search backend (dev)
---

# hub-search-v3 (dev backend)

Replacement backend for the Hub semantic-search Space. FastAPI + in-memory DuckDB
brute-force cosine search over Qwen3-Embedding-0.6B vectors. Drop-in compatible
with the old `davanstrien/huggingface-datasets-search-v2` API.

- Boot: loads `davanstrien/search-v2-embeddings` (both configs) into in-memory
  DuckDB `FLOAT[dim]` tables; loads the query model on CPU.
- Query encoding replicates the old backend so query vectors align with the seed docs.
- Endpoints: `/search/{datasets,models}`, `/similarity/{datasets,models}`,
  `/trending/{datasets,models}`, `/suggest/{type}`, `/health`.

Config via env: `EMB_DIM` (default 256, MRL truncation), `DUCKDB_THREADS`,
`EMBEDDINGS_REPO`. Needs `HF_TOKEN` secret for the private seed dataset.

Private dev Space — see the revival plan note for context.