@memharness/embed

@memharness/embed

Optional local embedding model for memharness hybrid (semantic) recall. Wraps BGE-small (Xenova/bge-small-en-v1.5, 384-dim, ~130MB) via transformers.js. The model is downloaded once from the HuggingFace hub, then runs fully offline: no API key, no per-query network.

This is a separate, optional package so that @memharness/core stays model-free and the default server install stays small. The @memharness/mcp server uses it only when MEMHARNESS_HYBRID=1 and this package is installed.

Use

import { embedDocuments, embedQuery, EMBED_MODEL, EMBED_DIM } from "@memharness/embed";

const [doc] = await embedDocuments(["user: drinks oolong tea"]); // store with setEmbedding
const q = await embedQuery("favorite beverage");                 // pass to recall({ queryVector })

embedQuery prepends the BGE retrieval instruction; embedDocuments does not. Pass setEmbedProgress(fn) before first use to surface model-download progress.

License

Apache-2.0