Skip to content

LIVE DEMO · ENTERPRISE DOCS RAG

Ask the documentation.

Live, cited answers from 202,234 Azure and Kubernetes passages — generated on one GPU and verified before you see them.

Enter to send · Shift+Enter for a new line

AZURE

KUBERNETES

The answer follows the language of the question.

AZURE PASSAGES

198,378

measured 2026-10-12

KUBERNETES PASSAGES

3,856

measured 2026-10-12

SOURCE PAGES

17,715

measured 2026-10-12

The seal locks only when every citation is a retrieved passage.

WHAT YOU JUST WATCHED

Five stages, timed end to end

01 ROUTE

Route

keyword guardrails + bge-m3 similarity · azure, kubernetes or both

02 RETRIEVE

Retrieve

bge-m3 · 1024-d · pgvector top-50

03 RERANK

Rerank

bge-m3 similarity blend · keep top-5

04 GENERATE

Generate

gemma-4-e4b · one RTX 3060 · answers only from the passages

05 VERIFY

Verify

every cited id must be a retrieved passage, or the answer is refused · a pass/fail gate — it has a result, not a duration

THE PROOF

What the seal means

Refuses instead of guessing

If no retrieved passage supports a claim, the gate returns a refusal, not a plausible answer.

Cites only what it retrieved

Every number in an answer points at a real documentation page you can open.

Runs on one consumer GPU

gemma-4-e4b runs on a single RTX 3060 behind the API. Questions and retrieved passages stay on our own hardware.

LIVErouting, retrieval, generation, verification, timings, citations

STATICcorpus counts (measured 2026-10-12) · model and GPU names · the index sweep is schematic

Refusals are a feature.

Built with

The Oracle — AI consulting & integration, from €1,500

The same pipeline — routing, retrieval, a self-hosted model and a citation gate — over your own documentation.

See the package

NEXT STEP

Want this over your documentation?

Private corpus, self-hosted model, verified answers — in weeks, not quarters.

Talk to usBack to demos