How the engagement works
1
Connect your data
Your corpus flows in through the connectors or the API — the same ingestion paths as the rest of the platform. Your data trains your model; it is not pooled into anyone else’s.
2
Freeze the evaluation set
We build a golden set from your real queries and judged outcomes — frozen before training starts, so the target can’t drift to flatter the model. This set is the acceptance gate for everything that follows.
3
Train and evaluate
The model is fine-tuned on your distribution and scored against the frozen set on the metrics that decide retrieval quality: recall on your queries, ranking position of the right answer, and calibrated no-answer behavior.
4
Promote only on a proven win
A candidate ships only when it beats the incumbent on your evaluation set. No vibes-based upgrades: if it doesn’t win on your queries, it doesn’t serve.
5
Serve pinned
The winning model serves your traffic with an explicit version pin — the same never-silently-substituted guarantee as the hosted API — alongside a reranker tuned to the same distribution. Ongoing maintenance keeps the model current as your data evolves.
What you end up with
- An embedding model that is measurably better on your queries than the generic API you use today — with the evaluation receipts to show it.
- A matched reranker, so first-stage recall and final ordering improve together.
- Serving in the same platform as your data — no cross-vendor pipeline to operate.
Start a conversation
Custom models are scoped per engagement — corpus size, query volume, and evaluation criteria shape the work.Book a call
30 minutes with the founder: bring a description of your corpus and a handful of real queries where retrieval disappoints you today.