One Zig binary that wants to be your search engine, vector DB, and ML stack
Antfly crams BM25, dense and sparse vectors, graph traversal, and in-process ML models into a single embeddable engine that scales from a file to a Raft cluster.

What it does
Antfly is a search-and-inference database written in Zig. One engine handles full-text (BM25), dense vectors (RaBitQ-compressed), sparse vectors (SPLADE), and graph traversal over the same table — and the models that chunk, embed, rerank, transcribe, and OCR run inside the process. Enrichment happens automatically as you write data: embeddings, chunks, entities, and graph edges get generated per index via configurable pipelines. The same engine deploys as a single .aflite file, a single node with hot standby, a multi-Raft cluster, or serverless over object storage (that last one is in progress).
The interesting bit
The scope is the story. Most stacks bolt together Postgres, a vector DB, an embedding service, and a reranker; Antfly wants all four jobs in one process, with a C API and WASM build so the whole thing can run in-process, in unit tests, or in the browser. There’s even a Postgres extension (antfly_postgres) exposing search via an @@@ operator, plus a psql-compatible wire protocol server.
Key highlights
- Hybrid search fusing BM25, dense, and sparse results via reciprocal rank or relative score fusion
- Built-in RAG agents with streaming, tool calling, web search, and TOON document rendering to cut prompt tokens
- Multimodal indexing of images, audio, and video in one vector space with EmbeddingGemma 2
- Native inference kernels: SIMD on x86/ARM, Metal on Apple silicon, CUDA with build-time-generated kernels
- Fine-tuning support (LoRA, QLoRA, SFT, DPO, GRPO) and bring-your-own models via Ollama, OpenAI, Bedrock, and others
- Postgres CDC mirroring, document TTL, ACID transactions, and Wasmtime-based extensions
Caveats
- The serverless-over-object-storage engine is explicitly marked in progress
- Automatic shard split/merge is opt-in, not default behavior
- The feature list is enormous; the README doesn’t say how mature each piece is, so expect uneven depth
Verdict
Worth a look if you’re assembling a RAG or hybrid-search stack and want fewer moving parts — especially appealing if you like the idea of the database embedding your documents literally and figuratively. If you need a battle-proven, single-purpose search engine or vector DB, this much ambition in a young project may give you pause.
Frequently asked
- What is antflydb/antfly?
- Antfly crams BM25, dense and sparse vectors, graph traversal, and in-process ML models into a single embeddable engine that scales from a file to a Raft cluster.
- Is antfly open source?
- Yes — antflydb/antfly is open source, released under the Apache-2.0 license.
- What language is antfly written in?
- antflydb/antfly is primarily written in Zig.
- How popular is antfly?
- antflydb/antfly has 517 stars on GitHub.
- Where can I find antfly?
- antflydb/antfly is on GitHub at https://github.com/antflydb/antfly.