Moorcheh on EdgeVector search that leaves room for the LLM.
moorcheh-edge compresses embeddings with MIB and searches them locally — text or precomputed vectors, optional on-device RAG via Ollama. No API keys. Proven on Arduino UNO Q at ~17 ms for 10k vectors.
Your Device. Your Knowledge.
One CLI, one Docker image, one local store — text or vectors, optional local RAG, no API keys.
One command. On-device.
moorcheh-edge up pulls moorcheh/moorcheh-edge, starts the API on localhost:8080, and warms BGE-small for text search. Add --with-llm for local RAG via Ollama (qwen2.5:0.5b-instruct). No API keys — Python 3.10+ and Docker.
Built for Local-First Vector Workflows
Same Moorcheh MIB/EDM engine. Sized for one device — search, optional RAG, and room for the LLM.
Nothing leaves the host
Vectors are binarized and stored on disk at ~/.moorcheh-edge/data. Queries are scored locally. Optional RAG calls Ollama on the same machine — no cloud vector DB.
No signup, no tokens
Run moorcheh-edge up and start uploading. There's no account, no billing, no rate limiter — just the local HTTP server on localhost:8080.
Information-theoretic codes
Maximum Information Binarization compresses float embeddings into compact one-bit codes scored with EDM — ~96 B per 768-d vector vs ~3 KB float32.
Tiny Docker footprint
On Arduino UNO Q we measured ~20 MiB container RSS at 10k vectors and ~17 ms median search — leaving RAM for BGE and a local LLM.
Text or bring your own
Text mode embeds with BGE-small (384-d) on the client. Vector mode accepts precomputed floats from 128–1536 dimensions. Same MIB store either way.
Works without the internet
After the first image and model pull, moorcheh-edge runs entirely offline. Built for air-gapped boxes, field devices, and kiosks like Arduino UNO Q.
Transparent Edge Quotas
What you get with the free moorcheh-edge store. Need more? Talk to us about Moorcheh.
One local store
Each moorcheh-edge store holds up to 10,000 items. Upload beyond the cap returns HTTP 409. Check usage via moorcheh-edge status.
Text or vector mode
Text mode locks to 384-d (BGE-small). Vector mode accepts 128, 256, 384, 768, 1024, or 1536 floats — set on first upload. Finite values only.
Search + local answer
Search works without an LLM. For moorcheh-edge answer, run up --with-llm to install Ollama and pull qwen2.5:0.5b-instruct on Linux.
{
"status": "ok",
"dimension": 384,
"store_mode": "text",
"embedding_model": "BAAI/bge-small-en-v1.5",
"items": 42,
"max_items": 10000,
"remaining": 9958
}