Skip to main content

Model reference — Documents

AI models for PII detection and redaction

This service finds personal data in text — names, addresses, telephone numbers, national identifiers, account numbers, dates of birth, medical details — and removes or masks it, so that records can be shared, analysed or used for training without exposing the people in them.

PII stands for personally identifiable information. Detection is only half the task: what happens next matters as much. A value can be blacked out, replaced by a label, replaced by a realistic but fictitious value, or replaced consistently by the same pseudonym everywhere it appears — the last being what allows records to stay linkable after redaction. Because a missed identifier is a reportable incident, these pipelines are tuned to over-flag and are measured on what they miss, not on their average accuracy.

PII detection and redaction service AI models for documents

Input type — Documents

Models in this group take text or whole documents as input: plain text, PDFs, scanned pages and office files. 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 one A4 page of about 500 words with all personal data found and masked.

Presidio

Vendor: Microsoft

What it does: a complete detection and masking framework: it combines pattern rules for structured identifiers such as card and insurance numbers with model-based detection for names and addresses, and applies your chosen masking. The usual backbone of a redaction pipeline.

RequirementMinimumMediumHigh
GPU typeNo GPU requiredRTX 4090A100 80 GB
VRAM24 GB80 GB
vCPUs4824
RAM8 GB32 GB64 GB
ServerCPU instance, 4 vCPU1× RTX 4090 24 GB1× A100 SXM 80 GB
Rate (pages/hour)≈ 5,000≈ 20,000≈ 60,000

GLiNER

Vendor: Urchade Zaratiana and contributors

What it does: finds any identifier type you name in plain words — "employee number", "patient identifier" — with no training data, which covers the identifiers specific to your business that no standard model knows.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 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)≈ 3,000≈ 12,000≈ 34,000

DeBERTa v3

Vendor: Microsoft

What it does: fitted to your own data from labelled examples, which gives the lowest miss rate on your document types. The right answer once a labelled sample exists.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 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)≈ 8,000≈ 30,000≈ 90,000

XLM-RoBERTa

Vendor: Meta

What it does: finds personal data across a hundred languages with one model, so an obligation is met uniformly rather than only in English.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 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≈ 24,000≈ 70,000

spaCy transformer pipelines

Vendor: Explosion

What it does: labels the standard entity types in a fast production pipeline, used as a broad first pass that a stricter model then reviews.

RequirementMinimumMediumHigh
GPU typeRTX 3060RTX 4090A100 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)≈ 12,000≈ 45,000≈ 130,000

Llama 3.1 8B

Vendor: Meta

What it does: catches personal data that only context reveals — a person identified by their role and location rather than by name — which pattern and small-model detection both miss.

RequirementMinimumMediumHigh
GPU typeRTX 3090RTX 4090H100 80 GB
VRAM16 GB24 GB80 GB
vCPUs81224
RAM32 GB48 GB128 GB
Server1× RTX 3090 24 GB1× RTX 4090 24 GB1× H100 SXM 80 GB
Rate (pages/hour)≈ 900≈ 2,800≈ 9,000

Choosing between them

What matters is your definition of personal data, the jurisdictions involved, and what the redacted text is used for. Our consultants review your data and obligations, then recommend the detection models, the masking strategy, and a measured miss rate you can put in front of an auditor.

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.

PII detection and redaction service AI models for documents Pricing

From benchmark to production

Share a representative sample, expected volume, latency target and deployment location for PII detection and redaction. 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.