/engine

Six stages, in order

The order is part of the story: a document becomes passages, passages become vectors, vectors become a ranked set, and only then does a model write anything.

Grounding Workflows Sovereignty
Step Stage What it does What we show
01

Ingest

Born-digital files, scans and photographed pages, processed one document at a time so structure and page images survive into the index.

one doc at a time OCR + page images
02

Chunk

Passages are cut at 700 tokens with a 120-token overlap, so a claim that straddles a boundary is still retrievable whole.

700 tokens 120 overlap
03

Embed

Passages are embedded with Voyage 3.5 at 1024 dimensions — the vector half of retrieval.

Voyage 3.5 1024 dims
04

Retrieve

Two independent methods run in parallel — BM25 on exact terms, vectors on meaning — fused with reciprocal rank fusion, re-scored by rerank-2.5, and cut to the top 8 passages.

hybrid BM25 + vector RRF fusion rerank-2.5 k=8
05

Generate

Generation is model-agnostic: the job selects the model, and the temperature is set per job rather than globally.

5 providers temperature per job
06

Govern

Every answer is bounded to the supplied corpus and carries citations. Tenants are separated by row-level security, airgapped deployment is supported, and changes are gated on evaluation.

closed corpus citations tenant RLS airgap eval
Generation providers
temperature set per job
Claude Anthropic
GPT OpenAI
Grok xAI
DeepSeek DeepSeek
Kimi Moonshot

Technical due diligence, under NDA

Configuration, evaluation scores and component ablation results are shared in a working session with your engineers, not published here.

Book a session