July 14, 2026
New in Numaga: embeddings through the gateway
Knowledge bases searchable by meaning, with cost and usage reported separately — inside the same governance as the rest of your AI traffic.
As of today, Numaga supports embeddings as a first-class part of the platform. Embeddings turn text into a numeric representation of meaning — the technique that lets a knowledge base find what you mean, even when your words differ from the document’s. Anyone working seriously with knowledge bases needs them; and precisely for that reason they belong inside your governance, not next to it.
Why this was needed
Knowledge bases grow fast. An organisation that connects thousands of documents — policy, manuals, registrations — quickly finds that keyword search falls short. Semantic search solves that, but until now it came with two problems: embedding traffic ran out of sight of governance, and the cost (real API rates, not token budgets) mixed with chat usage. That distorts every report.
So we routed embeddings through the same gateway as all other AI traffic. Same access policy, same logging, same cost controls — and vector storage sits in the managed database of your own environment, single-tenant in the Netherlands. No separate vector stack or external search service involved.
What you’ll notice
For users: better answers from the knowledge base, particularly in Dutch, and a combination of searching by meaning and by exact terms — the latter stays important for names, codes and domain jargon.
For administrators: a dedicated Embeddings overview in the dashboard. It shows which embedding models are active, how many calls run and what they cost — deliberately separated from the regular usage statistics, because putting tens of thousands of small embedding calls next to ordinary chat conversations is comparing apples and oranges.
Who this matters to
This is mainly relevant for organisations with large or growing knowledge bases. A concrete example from our own practice: a standard data source with tens of thousands of documents — including thousands of PDFs — is embedded end-to-end and stays searchable while the sync runs, without users noticing a thing.
Embeddings are on by default for new environments and can be enabled per environment for existing customers. Wondering what this means for your knowledge base? Book a demo — we’ll show it live.