Y0 Embeddings
availableThe index layer — meaning-preserving vectors tuned for the context graph.
[ 01 ]Spec sheet
Y0 Embeddings is the quiet family the rest of the platform stands on. Every document, message, transcript, and trace that enters the context graph is embedded on write, which is what makes retrieval feel like memory instead of search: when a run asks for 'the pricing discussion with the Bengaluru client', the graph resolves it by meaning, across files that never contain that exact phrase. The family ships two sizes — y0-embed-s at 512 dimensions for high-volume indexing where storage dominates cost, and y0-embed-l at 1536 dimensions where ranking quality on subtle distinctions earns its footprint. Both are trained with an emphasis the general-purpose alternatives lack: temporal and entity sensitivity, because in real work 'the contract' means the current contract, and the difference between invoice 41 and invoice 42 matters. Embeddings are also the engine behind deduplication, related-item surfacing, and the similarity joins that link a calendar event to the documents it is about. The API is deliberately boring — batch inputs, vectors out, stable across minor versions, with re-embedding handled by the platform on major version bumps so your index never silently mixes spaces. Generally available on every tier including free.
[ 02 ]Capabilities
Two sizes — 512d for volume, 1536d for ranking quality
Entity- and time-aware similarity tuned for working context
Automatic embed-on-write for everything entering the graph
Batch endpoint handling 2,048 inputs per call
Versioned vector spaces with managed re-embedding on upgrades
[ 03 ]Best for
Semantic retrieval over private corpora without running infra
Deduplication and clustering of operational documents
Hybrid search where vectors rank and filters constrain
[ 04 ]Sample request
{
"model": "y0-embed-l",
"input": [
"Renewal terms for the Meridian contract",
"Q3 invoice dispute - resolution summary"
],
"dimensions": 1536
}