Quanfire AI

We train small, sharp retrieval models for Indian languages and domains — and we can tell you where every training example came from. Provenance is not a footnote here; it is the product.

How we build

Refuse-by-default provenance. Every training source must carry an explicit licence basis — public-domain, owned, synthesized, or verified-open — before a single byte enters a corpus. Sources that wear a permissive label over content their publisher doesn't own are rejected, on principle, even when they'd help.
Honest scope on every card. We publish what a model does and, just as plainly, what it does not — including where transfer is flat. A retriever tuned for one distribution is described as exactly that.
Reproducible by construction. Every reported number carries the config fingerprint that produced it. Same fingerprint, same seed, same result.
Capital-efficient and from scratch. Trained on modest hardware, not a cluster. The moat is domain depth and clean data, not model size.

Published models

Model	What it is
`multilingual-embedding`	Flagship general-purpose multilingual retriever. One meaning searched across many languages.
`embed-legal-en`	English Supreme Court judgment retriever. +76% in-distribution Recall@1 (0.309→0.545, confidence intervals disjoint). Trained only on statutory public-domain judgment text; scope is English judgments, and the card says so.

Domain specialists follow the same discipline — legal and finance retrievers tuned for Indian text — and ship when the model earns its scope, not before.

Verify it yourself

Don't take the numbers on faith — run them. The public playground lets you search a single meaning across languages and watch a domain adapter separate the right answer from the noise:

playground.quanfire.ai