Node.js: Vector Embeddings

[sjvans]

No description provided.

[sjvans]

Note on the "lexical subword embeddings" wording (kept as-is per @MattSchur, flagging for discussion):

On SQLite/H2 the default vector_embedding emulation isn't a learned/semantic embedding — verifying against cds-dbs (sqlite/lib/SQLiteService.js, the VECTOR_EMBEDDING default), it's a deterministic hash: lowercase → extract 3-char n-grams → FNV-1a hash each → project into a 384-dim vector via the hashing trick → L2-normalize. The source comment itself calls it "hash-based," and the function is registered deterministic (it ignores text_type and model_and_version).

So "lexical subword embeddings" slightly oversells it: it captures character-n-gram overlap, not meaning. Two paraphrases with no shared trigrams score as unrelated. That's exactly why the doc then points to local ONNX models for real semantic embeddings — which is good — but readers may expect the default to already be semantic.

Proposal for a follow-up: call the default something like "a deterministic hash-based emulation (character n-grams)" and reserve "embeddings" for the ONNX path. Leaving the current wording in place for now; @MattSchur let's align.

[MattSchur]

"lexical subword embeddings" slightly oversells it: it captures character-n-gram overlap, not meaning

This is exactly lexical (or surface) vs. semantic embeddings, but I agree "embedding" might be confused with "captures semantics". What about "Lexical Character-Hash Vectors"?

[sjvans]

@MattSchur agreed — "Lexical Character-Hash Vectors" nails it: lexical (surface, not semantic), character-hash (the FNV-1a hash of char n-grams), and vectors rather than embeddings so it doesn't imply meaning. Applied to the SQLite/H2 note:

On SQLite and H2, the vector_embedding function is emulated using lexical character-hash vectors by default. These capture surface (character-n-gram) overlap, not meaning. To compute semantic embeddings, use local ONNX models.

Pushed to node-embeddings.