[MLsys2026 Best Paper]: https://arxiv.org/abs/2506.08276. RAG on Everything with LEANN. Enjoy 97% storage savings while running a fast, accurate, and 100% private RAG application on your personal device.
Python aifaissgpt-osslangchainllama-indexllmlocalstorageoffline-firstollamaprivacypythonragretrieval-augmented-generationvector-databasevector-searchvectors
docs: document LEANN_EMBEDDING_SERVER_HOST (#412)
Aakash Sureshb1fbec8871c1
diff --git a/docs/configuration-guide.md b/docs/configuration-guide.mdindex 06fc1fa..89268d6 100644--- a/docs/configuration-guide.md+++ b/docs/configuration-guide.md@@ -102,6 +102,30 @@ export LEANN_OLLAMA_HOST="http://localhost:11434" # falls back to OLLAMA_HOSTLEANN also recognises `LEANN_LOCAL_LLM_HOST` (highest priority), `LEANN_OPENAI_BASE_URL`, and `LOCAL_OPENAI_BASE_URL`, so existing scripts continue to work.+### Embedding Server Bind Address++The backend embedding servers (HNSW and DiskANN) listen on `127.0.0.1` by+default, so an index build or search does not expose an embedding endpoint to+the rest of the network. Override the bind address only when something outside+the machine genuinely needs to reach it — a container sidecar, or a searcher on+another host:++```bash+# Listen on every interface (containers, cross-host setups)+export LEANN_EMBEDDING_SERVER_HOST="0.0.0.0"++# Or bind one specific interface+export LEANN_EMBEDDING_SERVER_HOST="10.0.0.5"++# IPv6 works too; brackets are added automatically+export LEANN_EMBEDDING_SERVER_HOST="::1"+```++The server speaks an unauthenticated ZMQ REP protocol, so anything that can+reach the port can request embeddings. Prefer the loopback default, and if you+must widen it, restrict access at the firewall or container network rather than+exposing the port publicly.+### Passing Hosts Per Command```bash