Skip to main content
Most teams use a generic embedding API and hope it understands their domain. Conare’s position is different: the embedding model is part of the retrieval stack, not a commodity input — and the strongest version of it is one trained on your own data. Conare Embeddings comes in two layers:

Hosted API

POST /api/v1/embeddings — document-space vectors from the exact self-hosted model that indexes and retrieves Conare memories. Available today on every plan.

Custom models

An embedding model fine-tuned on your data, acceptance-gated on your real queries, served with its own reranker. Sales-led.

​
Why same-model vectors matter

Retrieval quality degrades silently when the vectors in your index and the vectors at query time come from different representations. The hosted API returns vectors from the same model Conare retrieves with, which makes it the right choice when:
  • you’re building your own vector features (clustering, dedup, similarity heuristics) next to Conare memory or ConareDB, and the representations must agree;
  • you want model pinning semantics — an explicit guarantee that vectors from a different model are never silently substituted into your index.

​
Why custom models exist

A generic model spreads its capacity across every domain on the internet. Your queries live in one domain — your product’s vocabulary, your customers’ phrasing, your documents’ structure. Fine-tuning on that distribution reliably moves the metrics that matter (recall on your real queries, ranking quality, knowing when there is no answer) — and Conare only ships a custom model when it provably beats the incumbent on your own evaluation set, never on faith. Read how the engagement works.