All sectors

Routing document pages before ingestion into an assistant

Each page of a PDF document is automatically routed either to free text extraction or to reading by a vision model, only when that is useful. The client ingests faster and pays for costly reading only on the pages that need it.

Talk to an engineer

Illustrative mock-up: the names, values and documents shown are fictitious.

Context

An organisation feeds its AI assistant with internal documents synchronised daily from SharePoint: course materials, exams, procedures, charters.

The problem

Sending every page to a vision model is expensive and slows ingestion down; relying on native extraction alone loses figures, formulas and scanned pages.

What we built

A page-by-page router: a frozen visual encoder and three small classifiers predict whether native extraction will fail (useful image, formula, scan), plus two model-free text pre-filters. It runs on the GPU already used by the search pipeline, with no dedicated hardware.

Steps

01Building a corpus of 20 documents, 999 pages and 1,951 verification questions
02Training three decision heads on a frozen visual encoder
03Measuring the gain at equal cost, validated by a second annotator and replicated on a public benchmark
04Integration into the client's ingestion pipeline

Hosting and models

Frozen DINOv3 encoder (SigLIP 2 or DINOv2 possible, under the Apache 2.0 licence), inference on a T4 GPU shared with the embedder and the reranker. Vision model called for routed pages (Qwen3.5-2B and Gemini Flash evaluated).

Services involved

AI in your applications · Forward Deployed Engineers

We name a client only with their written agreement. The budgets and detailed results of our engagements remain confidential.