Skip to main content

Model reference — Documents

AI models for Document question answering

Document question answering lets someone ask a question in ordinary language — “what notice period applies to the Frankfurt lease?” — and get the answer back with the passage and page it came from, instead of searching a folder of PDFs by hand.

Two kinds of model are needed. An embedding model turns every passage in your archive into a numeric fingerprint so the passages closest in meaning to a question can be found; the same is done to the question at the moment it is asked. A language model then reads the retrieved passages and writes the answer, quoting its source. A third, optional model — a reranker — re-sorts the retrieved passages by relevance before the language model sees them, which measurably improves the answers on large archives.

Document question answering service AI models for documents

Input type — Documents

Models in this group take documents and questions as input: PDFs, scanned pages, office files and the plain text extracted from them. Each specification table gives three hardware tiers — Minimum, the smallest setup on which the model runs correctly; Medium, the usual production configuration; and High, a configuration sized for peak volume. The rate is the number of sample inputs processed per hour on that hardware. Use these rates for initial sizing. Before production, benchmark your own data to validate accuracy, latency, throughput and cost. The sample input here is stated per model — one A4 page of about 500 words for indexing models, or one question answered from retrieved passages for answering models.

BGE-M3

Vendor: Beijing Academy of Artificial Intelligence

What it does: turns passages and questions into numeric fingerprints so that passages matching a question’s meaning can be found even when they share no words with it. It handles over a hundred languages and passages up to about 8,000 words, which suits long contracts and reports.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 40901× A100 80 GB
VRAM8 GB24 GB80 GB
vCPUs61224
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 6,000≈ 30,000≈ 90,000

E5-large / multilingual-E5

Vendor: Microsoft

What it does: a well-proven fingerprinting model for retrieval, small and quick to run. A dependable default for English archives, with a multilingual version for mixed-language collections.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 40901× A100 80 GB
VRAM6 GB24 GB80 GB
vCPUs4824
RAM16 GB32 GB64 GB
Server1× RTX 3060 12 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 9,000≈ 45,000≈ 140,000

BGE Reranker v2-M3

Vendor: Beijing Academy of Artificial Intelligence

What it does: reads a question and a candidate passage together and scores how well the passage actually answers it. Run over the top few dozen candidates, it pushes the genuinely relevant passage to the top, which is often the single biggest accuracy gain in the pipeline.

RequirementMinimumMediumHigh
GPU typeRTX 3090RTX 40901× A100 80 GB
VRAM12 GB24 GB80 GB
vCPUs81224
RAM32 GB48 GB96 GB
Server1× RTX 3090 24 GB1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (question-passage pairs/hour)≈ 40,000≈ 150,000≈ 400,000

Llama 3.3 70B

Vendor: Meta

What it does: reads the retrieved passages and writes the answer with citations. The larger size is worth its cost when questions require combining facts from several documents rather than quoting one.

RequirementMinimumMediumHigh
GPU type2× RTX 4090 (reduced precision)1× H100 80 GB4× H100 80 GB
VRAM48 GB combined80 GB320 GB combined
vCPUs162464
RAM64 GB128 GB512 GB
Server2× RTX 4090 24 GB1× H100 SXM 80 GB4× H100 SXM 80 GB
Rate (answers/hour)≈ 120≈ 600≈ 3,000

Qwen2.5 32B

Vendor: Alibaba Cloud

What it does: a mid-sized answering model that handles very long retrieved contexts, so it can be given a whole chapter rather than a few paragraphs. Strong across European and Asian languages.

RequirementMinimumMediumHigh
GPU typeRTX 4090 (reduced precision)1× L40S 48 GB2× H100 80 GB
VRAM22 GB48 GB160 GB combined
vCPUs121648
RAM48 GB64 GB256 GB
Server1× RTX 4090 24 GB1× L40S 48 GB2× H100 SXM 80 GB
Rate (answers/hour)≈ 180≈ 560≈ 2,200

Mistral Small 3

Vendor: Mistral AI

What it does: a compact answering model that fits on one card and keeps latency low, which matters when staff expect a reply while they are still on the phone to a customer.

RequirementMinimumMediumHigh
GPU typeRTX 40901× L40S 48 GB1× H100 80 GB
VRAM24 GB48 GB80 GB
vCPUs121632
RAM48 GB64 GB128 GB
Server1× RTX 4090 24 GB1× L40S 48 GB1× H100 SXM 80 GB
Rate (answers/hour)≈ 260≈ 700≈ 2,200

ColPali

Vendor: Illuin Technology

What it does: fingerprints the page image itself rather than extracted text, so it retrieves the right page from scans, slides and charts where text extraction would lose the layout. Useful for archives full of presentations and forms.

RequirementMinimumMediumHigh
GPU typeRTX 40901× A100 80 GB2× A100 80 GB
VRAM20 GB80 GB160 GB combined
vCPUs81632
RAM32 GB64 GB128 GB
Server1× RTX 4090 24 GB1× A100 SXM 80 GB2× A100 SXM 80 GB
Rate (pages/hour)≈ 900≈ 3,500≈ 7,000

Choosing between them

The choice is really a choice of pipeline: how your archive is indexed, whether a reranker earns its keep at your volume, and how large the answering model needs to be. Our consultants review the archive, the questions your staff actually ask and the accuracy you need on citations, then recommend the combination that fits.

At the start of a project we may run a short proof of concept on a sample of your own data, measuring the accuracy and the throughput the model actually achieves on your material. That replaces the estimates on this page with real figures, so the cost and the schedule for the full engagement are known before it is committed.

Document question answering service AI models for documents Pricing

From benchmark to production

Share a representative sample, expected volume, latency target and deployment location for Document question answering. We benchmark the shortlisted models, recommend the lowest-cost GPU configuration that meets the target, and scale it from pilot capacity to a dedicated production cluster.