extract
LightOn's extract () is a document intelligence API that pulls specific fields, entities, or key-value pairs from documents and returns structured JSON.
Publisher review
LightOn's extract () is a document intelligence API that pulls specific fields, entities, or key-value pairs from documents and returns structured JSON. It targets enterprise teams—particularly those in regulated industries like finance, insurance, and government—who need to automate data extraction from invoices, contracts, lease agreements, and claims forms without exposing sensitive data to public cloud inference. The tool is part of LightOn’s broader platform, which includes parsing (LightOnOCR-2) and grounded retrieval (/search), all accessible via a single API key on LightOn Console or deployed on-premises for digital sovereignty.
To use extract (), you define a JSON schema specifying the fields you care about—e.g., invoice number, lease end date, claim ID, or contract clause—and submit a document via the /extract endpoint. The API processes the document asynchronously, returns results via webhooks, and cites each extracted field back to its source passage in the document. Pricing is €0.004 per page for extraction, with no additional per-field charge. The underlying engine is LightOnOCR-2, which achieves state-of-the-art 83.2 on OLMOCR-BENCH and natively supports 20+ languages, including handwriting and multi-column layouts. LightOn also offers /parse (€0.002/page) and /search (€0.006/query) endpoints, with hybrid retrieval using dense, sparse, and late-interaction signals, and sub-200ms P50 latency.
LightOn was founded in 2016, is headquartered in Paris, France, and became Europe’s first publicly listed generative AI startup in 2024. It competes with Emergent Risk International, AmpiFire, Kodehash Technologies, Reggora, and Vassar Digital—most of which are smaller or less focused on sovereign document AI. LightOn differentiates through its proprietary Optical Processing Unit (OPU) hardware, open-source retrieval models (LateOn, NextPlaid, PyLate) with 50M+ HuggingFace downloads and 916K monthly PYPI installs, and partnerships with Orange Business and Hewlett Packard Enterprise for on-premises deployment. The platform supports MCP-native integration, LLM-agnostic retrieval, and chunk-level ACLs for multi-agent systems.
The honest trade-offs: First, LightOn has limited customer reviews and detailed testimonials, making it harder to validate real-world performance beyond benchmark scores. Second, pricing is per-page and per-query with no published volume discounts or enterprise tiers, which could surprise high-volume users. Third, the tool is optimized for structured extraction from documents—it is not a general-purpose OCR or data pipeline tool for unstructured video or audio. Fourth, while on-premises deployment ensures data sovereignty, it requires infrastructure investment and LightOn’s proprietary OPU hardware for optimal acceleration, which may not fit every IT budget.
How it works
-
Schema-defined field extraction
Users define a JSON schema for fields like invoice numbers or claim IDs; the API returns structured JSON with citations per field.
-
Async processing with webhooks
Extraction runs asynchronously; results are delivered via webhooks, enabling non-blocking workflows for large document batches.
-
Grounded citations per field
Every extracted value is cited back to its source passage in the document, enabling auditability and trust in the output.
-
LightOnOCR-2 parsing engine
State-of-the-art OCR with 83.2 OLMOCR-BENCH score, handling scans, tables, handwriting, and multi-column layouts in 20+ languages.
-
Hybrid retrieval integration
The /search endpoint uses dense, sparse, and late-interaction signals; the index selects the optimal signal per query.
-
On-premises and sovereign deployment
Services can be deployed air-gapped, hybrid, or as sovereign SaaS, protecting intellectual property and sensitive data.
-
MCP-native agent integration
Drop LightOn into any agent speaking Model Context Protocol; supports single-agent, multi-agent, and business application workflows.
Strengths and trade-offs
Strengths
- Achieves state-of-the-art 83.2 on OLMOCR-BENCH for document parsing, outperforming many general-purpose OCR engines.
- Pricing is transparent and usage-based at €0.004 per page for extraction, with no hidden per-field fees.
- Supports 20+ languages natively, including handwriting and complex multi-column layouts, reducing pre-processing needs.
- Open-source retrieval models (LateOn, NextPlaid) have 50M+ HuggingFace downloads and 916K monthly PYPI installs, ensuring community-tested reliability.
Trade-offs
- Limited customer reviews and detailed testimonials make it difficult to assess real-world performance outside benchmark scores.
- No published volume discounts or enterprise pricing tiers, which could increase costs for high-volume extraction workflows.
- Optimized for structured document extraction, not suitable for unstructured video, audio, or real-time streaming data.
- On-premises deployment requires infrastructure investment and proprietary OPU hardware for optimal acceleration, adding upfront cost.
Pricing context
€0.004 per page for extraction (extract endpoint); €0.002 per page for parsing (parse endpoint); €0.006 per query for retrieval (search endpoint). No published volume discounts or enterprise tiers.
Getting started with extract
-
Sign up for LightOn Console
Go to the LightOn Console website and create an account. Provide your email and set a password. Verify your email to activate the account. This gives you access to the API key and dashboard for managing your extraction projects.
-
Generate your API key
Log in to LightOn Console and navigate to the API keys section. Click "Create new key" and copy the generated key. Store it securely; you will use it to authenticate all requests to the /extract endpoint.
-
Define extraction schema
Create a JSON schema that lists the fields you want to extract, such as invoice number or lease end date. Specify the data type and optional constraints for each field. This schema tells the API exactly what to pull from your documents.
-
Submit a document for extraction
Send a POST request to the /extract endpoint with your API key, the document file (PDF or image), and the JSON schema. The API processes the document asynchronously and returns a job ID. Set up a webhook URL to receive the results when processing completes.
-
Review extracted data with citations
When the webhook delivers the results, inspect the structured JSON output. Each extracted field includes a citation pointing to its source passage in the document. Verify the accuracy of the data and adjust your schema if needed for future extractions.
Frequently Asked Questions
What is LightOn extract and how does it work for document extraction?
LightOn extract is a document intelligence API that pulls specific fields, entities, or key-value pairs from documents and returns structured JSON. You define a JSON schema for fields like invoice numbers, submit a document via the /extract endpoint, and get results asynchronously via webhooks.
How much does LightOn extract cost per page?
LightOn extract costs €0.004 per page for extraction, with no additional per-field charges. There is also a /parse endpoint at €0.002 per page and a /search endpoint at €0.006 per query. No published volume discounts or enterprise tiers are available.
What types of documents can LightOn extract handle?
LightOn extract handles invoices, contracts, lease agreements, and claims forms. It uses LightOnOCR-2, which achieves a state-of-the-art 83.2 on OLMOCR-BENCH and supports 20+ languages, including handwriting and multi-column layouts. It is optimized for structured document extraction, not unstructured video or audio.
Can LightOn extract be deployed on-premises for data sovereignty?
Yes, LightOn extract can be deployed on-premises, air-gapped, or as sovereign SaaS to protect sensitive data. This requires infrastructure investment and LightOn’s proprietary Optical Processing Unit hardware for optimal acceleration. It targets regulated industries like finance, insurance, and government.
Does LightOn extract provide citations for extracted fields?
Yes, every extracted value is cited back to its source passage in the document, enabling auditability and trust. This grounded citation feature helps enterprise teams verify outputs and ensures compliance in regulated workflows where data accuracy is critical.
How does LightOn extract integrate with AI agents and workflows?
LightOn extract supports MCP-native integration, allowing it to drop into any agent speaking Model Context Protocol. It works with single-agent, multi-agent, and business application workflows, and supports chunk-level ACLs for multi-agent systems. It also offers hybrid retrieval via the /search endpoint.
Alternatives
- Octoparse ↗
- Apify ↗
- ScraperAPI ↗
How extract compares
Direct head-to-head against 3 competitors. Picked by 7wData.
extract
- Pricing
- €0.004 per page for extraction (extract endpoint); €0.002 per page for parsing (parse endpoint); €0.006 per query for retrieval (search endpoint). No published volume discounts or enterprise tiers.
- Target
- LightOn's extract () is a document intelligence API that pulls specific fields, entities, or key-value pairs from documents and returns structured JSON.
- Strength
- Achieves state-of-the-art 83.2 on OLMOCR-BENCH for document parsing, outperforming many general-purpose OCR engines.
- Watch for
- Limited customer reviews and detailed testimonials make it difficult to assess real-world performance outside benchmark scores.
Octoparse
- Pricing
- Standard: $89/month
- Target
- Marketing teams needing visual workflows
- Deployment
- Cloud & desktop
- Strength
- Prebuilt templates for common sites
- Watch for
- Cloud costs escalate with scale
Apify
- Pricing
- Starter: $49/month
- Target
- Developers building scrapers
- Deployment
- Cloud
- Strength
- Headless browser automation
- Watch for
- Steep learning curve for non-coders
ScraperAPI
- Pricing
- Starter: $29/month
- Target
- APIs for production apps
- Deployment
- Cloud
- Strength
- Proxy rotation & CAPTCHA solving
- Watch for
- Per-request pricing adds up
User reviews
No user reviews yet. Be the first to write one.
Sources
Reporting on this tool draws on these publicly available sources.