Back to Showcase
Honeycomb
Semantic search for 100K+ animated GIFs, built in a weekend with Antfly Swarm and Termite embeddings.
Architecture
1
Source Data
102K GIFs from the GIFDB dataset
2
AI Description
Gemini 2.0 Flash Lite generates text descriptions for each GIF
3
Embed with Termite
bge-small-en-v1.5 encodes descriptions into vectors locally
4
Index with Antfly
Hybrid BM25 + vector index stores everything in a single database
5
Search
Queries hit both BM25 and vector indexes, results are fused and ranked
The Project
Honeycomb indexes over 102,000 animated GIFs with semantic search powered by Antfly and Termite. Instead of relying on tags or filenames, users search by describing what they want — "cat falling off table" or "excited celebration" — and get semantically relevant results.
Stack
- Antfly (database + hybrid search)
- Termite (local embeddings with bge-small-en-v1.5)
- Gemini 2.0 Flash Lite (GIF descriptions)
- React + TypeScript + Tailwind + Vite (frontend)
Takeaways
The combination of Antfly's hybrid search and Termite's local inference meant zero external API calls for the search pipeline. The entire embedding and indexing pipeline runs locally, making it fast and free to iterate.
Swarm mode was the killer feature — I had a working search engine in under 5 minutes.