“We were scaling document volume by nearly 50 percent every month —
Nutrient handled it without us adding resources.”
Harvey
compare
Nutrient is an Amazon Textract alternative for document extraction and retrieval-augmented generation (RAG). This feature-by-feature comparison shows where Nutrient wins — deploy-anywhere extraction, 100+ languages, and LLM-ready grounded output — and where Amazon Textract has the edge.
| Nutrient | Amazon Textract | |
|---|---|---|
| Core approach | An owned, hybrid optical character recognition (OCR) + AI
extraction pipeline, offered as a cloud API and as self-hosted
Java and Python SDKs. Tunable modes return structured Markdown,
spatial JSON, or schema-shaped JSON. | An Amazon Web Services (AWS) machine learning document service.
General OCR plus feature APIs (forms, tables, layout, queries,
signatures) and vertical models, returning a directed block graph. |
| Deployment | ||
| Output format | ||
| Language coverage |
Nutrient runs extraction as a hosted cloud API, or as self-hosted Java and Python SDKs that process documents on your own infrastructure — fully offline and air-gapped when required, so sensitive documents never leave your environment.
Textract runs only inside AWS. There is no self-hosted or on-premises option, and every document is processed in the AWS cloud. Moving an AWS-native Textract pipeline elsewhere means rearchitecting around S3, Lambda, and SNS/SQS.
Nutrient returns whole-document Markdown and typed spatial elements with reading order preserved — structure a RAG pipeline or LLM can consume directly, plus schema-shaped JSON with per-field citations.
Textract returns a directed block graph of pages, lines, words, and key-value sets. Turning that into reading-ordered, LLM-ready text is left to your code or AWS’s helper libraries. On the independent ParseBench, Textract’s semantic-formatting score was 3.7 percent.
Nutrient publishes an open, reproducible extraction benchmark on the public opendataloader-bench corpus — 200 hand-annotated PDFs, rerun on every release. Amazon Textract publishes no first-party accuracy benchmark; its only public number comes from an independent third party.
| OpenDataLoader benchmark | Nutrient |
|---|---|
| Overall accuracy (understand mode) | 0.93 |
| Reading order (NID) | 0.96 |
| Table structure (TEDS, understand mode) | 0.94 |
| Speed per page (default engine) | 0.004s — 134× faster than Docling |
Nutrient scores from the public opendataloader-bench corpus (200 PDFs, hand-annotated ground truth, three metrics: reading order, table structure, heading hierarchy), rerun on every release. Nutrient’s open source tool is listed on the opendataloader.org leaderboard; vision modes are run internally against the same corpus. Full results are on the Data Extraction API benchmarks page, with the methodology in the opendataloader-bench writeup. Nutrient also publishes an open grounding benchmark (grounding-en, Apache-2.0) that ranks #1–#2 on the public leaderboard. Amazon Textract isn’t evaluated on opendataloader-bench — these are Nutrient’s own published results, not a head-to-head.
On the independent ParseBench evaluation of extraction APIs (2,000 human-verified pages), Amazon Textract scored 47.9 percent overall — the lowest of the major cloud OCR services tested. Its table extraction was strong (84.6 percent), but semantic formatting (3.7 percent) and chart understanding (6.0 percent) were near the bottom of the field, reflecting a raw-extraction engine rather than a document-understanding one. Nutrient wasn’t part of the ParseBench cohort, and the two benchmarks use different corpora, so these figures aren’t a head-to-head — they’re each vendor’s standing on a separate public test.
An honest, extraction-by-extraction comparison — including where Textract wins.
| Nutrient | Amazon Textract | Winner | |
|---|---|---|---|
| OCR and language coverage | Printed-text OCR in 100+ languages, including CJK, Arabic, and
Cyrillic scripts. | Printed text in six Latin-script languages; advanced features are
English-only. | Nutrient |
| Table extraction | Row, column, and cell structure with spans; 0.94 TEDS in
understand mode on opendataloader-bench. | A core strength — structured and semi-structured tables; 84.6
percent on the independent ParseBench. | Both strong |
| Forms/key-value pairs | Key-value regions with relationship confidence, in understand
mode, across 100+ languages. | Mature forms feature, but English only and billed at a premium —
about $50 per 1,000 pages. | Nutrient |
| Layout and reading order | Reading order preserved into Markdown; 0.96 NID on
opendataloader-bench. | Layout feature classifies blocks, but reading-order linearization
is left to the caller. | Nutrient |
| Targeted field extraction | The /extract endpoint maps a document to your JSON Schema
with per-field citations. | Natural-language queries (English), plus adapters to fine-tune
them. | Different approaches |
| Prebuilt vertical APIs (invoices, IDs, lending) | Schema-driven — any document type via a JSON Schema you define. | Turnkey AnalyzeExpense, AnalyzeID, and AnalyzeLending with
built-in field taxonomies. | Amazon Textract |
| Handwriting | Handwritten regions recognized on your own infrastructure via the
vision tier. | Handwriting detection in the cloud — English only. | Draw |
| Signature detection | Not a dedicated extraction feature. | Dedicated signatures feature detects handwritten signatures and
initials. | Amazon Textract |
| Grounding and citations | Interpretable match labels (exact, fuzzy, not found) and source
blocks per field, plus a published grounding benchmark. | Per-block confidence scores and bounding boxes; no interpretable
grounding label. | Nutrient |
| LLM-ready output | Whole-document Markdown, spatial JSON, and schema JSON from one
API. | A block graph the caller traverses and linearizes. | Nutrient |
| Self-hosted/air-gapped extraction | Run extraction on your own infrastructure with the Java and
Python SDKs, offline when required. | AWS cloud only. | Nutrient |
| Asynchronous batch at scale | The cloud API is synchronous; self-hosted SDKs scale across your
own workers. | Asynchronous jobs process up to 3,000 pages/500 MB per document. | Amazon Textract |
| Human-in-the-loop review | Per-field confidence and match labels route low-confidence fields
to human review in your own workflow. | Amazon Augmented AI (A2I) provides a managed human-review
workflow with a built-in reviewer console. | Draw |
| Open, reproducible benchmark | Public opendataloader-bench results, rerun every release; listed
on opendataloader.org. | No first-party public accuracy benchmark. | Nutrient |
| Price for structured extraction | One understand-mode call returns layout, tables, and key-value
pairs together. | Features are billed additively — forms + tables + queries reaches
$70 per 1,000 pages. | Nutrient |
Textract’s base OCR is inexpensive, but its extraction features are billed additively — turning on forms, or forms plus tables plus queries, multiplies the per-page cost. Nutrient bills per-page credits by processing mode, and a single understand-mode call returns layout, tables, and key-value pairs together. Here’s a rough guide for 1,000 pages.
| Nutrient | Amazon Textract | |
|---|---|---|
| Free tier | 5,000 credits/month | 1,000 OCR pages/month (first 3 months); 100 pages/month for
Forms, tables, or queries |
| Text/OCR only | ~$0.84–$2.00/1,000 pages | ~$1.50/1,000 pages |
| Tables | ~$8–$18/1,000 pages (understand) | ~$15/1,000 pages |
| Forms/key-value | ~$8–$18/1,000 pages (understand) | ~$50/1,000 pages |
| Forms + tables + queries | ~$8–$18/1,000 pages (one mode) | ~$70/1,000 pages |
Approximate self-serve rates as of July 2026. Amazon Textract US-East, first 1M pages/month: DetectDocumentText $1.50, Tables $15, Forms $50, and Forms + Tables + Queries $70 per 1,000 pages — features are billed additively. Nutrient bills per-page credits by processing mode (text, structure, understand, agentic), with 5,000 free credits every month; a single understand-mode call returns layout, tables, and key-value pairs together. Textract’s base OCR is inexpensive; costs rise sharply once structured extraction features are enabled.
Run extraction as a hosted cloud API or entirely self-hosted with the Java and Python SDKs — offline and air-gapped when required, on any cloud or on-premises. Textract runs only in AWS.
Whole-document Markdown, spatial JSON, or schema-shaped JSON with reading order preserved — not a block graph you have to traverse and linearize.
OCR in 100+ languages, including CJK, Arabic, and Cyrillic. Textract handles six Latin-script languages, and its advanced features are English-only.
Per-field citations and interpretable match labels, backed by a public opendataloader-bench score rerun on every release and an open grounding benchmark.
5,000 Data Extraction API credits every month, no credit card required, and we don’t train on your documents.
Amazon Textract is an AWS machine learning service for optical character recognition (OCR) and document data extraction. It offers general text detection; an AnalyzeDocument API with features for forms, tables, layout, queries, and signatures; and vertical models (AnalyzeExpense for invoices, AnalyzeID for identity documents, AnalyzeLending for mortgages). It runs only inside AWS and returns a directed graph of block objects.
Yes, especially for teams that need to extract outside AWS, self-host or air-gap the pipeline; process documents in more than a handful of languages; or feed an LLM directly. Nutrient offers a cloud API and self-hosted Java and Python SDKs, 100+ language OCR, LLM-ready Markdown and schema JSON, and grounded output with a published accuracy benchmark. Textract has the edge on turnkey vertical APIs, native human-in-the-loop review (A2I), and very large asynchronous batch jobs inside AWS.
The main alternatives are Nutrient, Azure Document Intelligence, Google Document AI, LlamaParse, Reducto, Unstructured, and Docling. Nutrient stands out for deploy-anywhere extraction (cloud or fully self-hosted), 100+ language coverage, LLM-ready structured output, and grounded accuracy published on a public, reproducible benchmark.
Nutrient returns reading-ordered Markdown and schema-shaped JSON with per-field citations — structure a retrieval-augmented generation (RAG) pipeline can chunk and an LLM can consume directly. Textract returns a block graph that the caller must traverse and linearize; on the independent ParseBench, its semantic-formatting score was 3.7 percent. For RAG ingestion, Nutrient’s output is closer to LLM-ready out of the box.
Yes, with Nutrient. Beyond the hosted API, Nutrient’s extraction runs through self-hosted Java and Python SDKs and Document Engine, so documents can be processed inside your own infrastructure — including offline and air-gapped deployments. Amazon Textract only runs in the AWS cloud, with no self-hosted or on-premises option.
Yes. Nutrient performs OCR in 100+ languages, including Chinese, Japanese, Korean, Arabic, and Cyrillic scripts. Amazon Textract detects printed text in six Latin-script languages (English, French, German, Italian, Portuguese, Spanish), and its handwriting, queries, and vertical APIs are English-only.
Textract’s base OCR (DetectDocumentText) is inexpensive at about $1.50 per 1,000 pages, but structured extraction is billed per feature and stacks: Forms is about $50 per 1,000 pages, and forms plus tables plus queries reaches about $70. Nutrient bills per-page credits by processing mode, and one understand-mode call returns layout, tables, and key-value pairs together, with 5,000 free credits every month. Talk to our team for a comparison scoped to your document mix.
AWS now recommends Bedrock Data Automation as the starting point for most new intelligent document processing projects, positioning Textract as a lower-level OCR and table primitive. Teams choosing a long-term extraction layer should weigh that direction — and whether a single vendor that runs in any cloud or on-premises is a more durable foundation than an AWS-only primitive.
EXPLORE
Reducto is a strong agentic document extraction platform with state-of-the-art table parsing. Nutrient is the broader, deterministic document platform — extraction plus viewing, editing, signing, and conversion — at a fraction of the per-page cost.
Unstructured.io is a strong RAG-ingestion toolkit — open source partitioning, chunking, and a deep connector ecosystem. Nutrient adds what it doesn’t: grounded schema extraction and the full document lifecycle — viewing, editing, signing, and conversion.
LlamaIndex’s LlamaParse and LlamaExtract are cloud-first — self-hosted BYOC is gated to Enterprise plans — and lean on foundation model inference. Nutrient delivers deterministic, source-grounded extraction, self-hosted on any plan, and a viewer to verify every citation.
5,000 free Data Extraction API credits per month — no credit card required. Parse and extract source-grounded data your AI workflows can trust, in the cloud or on your own infrastructure.