verb · → vectors

embed

Turn text into vectors that sit close together when meanings are close. Use them for semantic search, retrieval for RAG, deduplication and clustering.

One or many#

import { embed } from 'edgewise';

const { embedding, dimensions } = await embed({ model: 'embed:default', input: 'How do I reset my password?' });

const { embeddings } = await embed({
  model: 'embed:tiny',
  values: docs.map((d) => d.text),
  onBatch: ({ done, total }) => (bar.value = done / total),
});

input takes one string and returns embedding. values takes many and returns embeddings, in batches of 16. Vectors are Float32Array, normalized to unit length unless you pass normalize: false, so a dot product is the cosine similarity.

Queries and documents#

EmbeddingGemma embeds search queries and documents with different prefixes. Say which one you are embedding and Edgewise adds the right prefix:

await embed({ model: 'embeddinggemma-300m', values: chunks, purpose: 'document' });
await embed({ model: 'embeddinggemma-300m', input: question, purpose: 'query' });

Models without prefixes ignore purpose.

Shorter vectors#

EmbeddingGemma is trained so its first dimensions carry the most meaning. Pass dimensions (768, 512, 256 or 128) to trade a little accuracy for a smaller index:

const r = await embed({ model: 'embeddinggemma-300m', input: 'hello', dimensions: 256 });
r.model; // 'embeddinggemma-300m@256'

result.model records the model and size. Store it with your vectors: vectors from different models, or different sizes, cannot be compared.

Similarity#

import { cosine } from 'edgewise/helpers';
cosine(a, b); // -1..1

For search over many vectors use vectorIndex().

Models#