LightOn API - Parse, Extract, and Search Documents

In production at

Three endpoints.
Zero pipeline maintenance.

Parse a document. Extract any field. Retrieve with citations.
The same API key on Console, the same SDK on Enterprise.

LightOnOCR-2.
State-of-the-art parsing.

Turns scans, tables, handwriting, and multi-column layouts into structured Markdown. 20+ languages natively. The parsing engine behind every retrieval workflow.

83.2

OLMOCR-BENCH (SOTA)

€0.002

Per Page

20+

Native language

Open

Weights on HuggingFace

$ curl https://api.lighton.ai/v1/parse \    
    -H 'Authorization: Bearer $LIGHTON_API_KEY' \ 
    -H 'Content-Type: application/json'\ 
    -D '{"document":"https://console-examples.lighton.ai/AFD-091005-062.pdf"}'

Define the schema.
Get JSON back.

Pull any field, entity, or key-value pair you care about. Invoice numbers, lease end dates, claim IDs, contract clauses. You define the schema; LightOn returns structured JSON.

JSON Schema

In / Out

€0.004

Per Page

Async

Via Webhooks

Cited

Per Field

$ curl https://api.lighton.ai/v1/extract \    
    -H 'Authorization: Bearer $LIGHTON_API_KEY' \   
    -H 'Content-Type: application/json' \   
    -d ''{"document":"https://console-examples.lighton.ai/AFD-091005-062.pdf","schema":"<YOUR_JSON_SCHEMA>"}'

Grounded retrieval
with citations.

One query, three signals: dense, sparse, late-interaction. The index picks the right signal, not the developer. Every result ships with the source passage that produced it. Built on LateOn and NextPlaid, our open-source ColBERT family.

Multi-vector

Dense + Sparse + Li

€0.006

Per Query

<200ms

P50 Latency

ACL

At Chunk Level

$ curl https://api.lighton.ai/v1/search \
    -H 'Authorization: Bearer $LIGHTON_API_KEY' \
    -H 'Content-Type: application/json' \
    -d '{"query":"<YOUR_QUERY>"}'

Try it in your browser.
No install needed.

Every endpoint is testable directly from Console.
Drop a file, get the response, copy the code into your project.

Open the playground

console.lighton.ai

Start free.
‍Pay per page and per query.

No seat licenses. No commitment. The free tier covers most prototypes.
Per-page pricing kicks in only when you scale.

STARTER

For developers and builders.
‍

Free

+ PAYG

Pure usage-based — no commitment.
‍

Start for free

Read the docs

Includes

Unlimited workspaces

Native data source connectors

Data governance & access controls (RBAC)

SSO (SAML, OIDC, LDAP) & SCIM provisioning

Audit logs & usage analytics

Bring Your Own LLM or LightOn LLMaaS

EU sovereign hosting

BUSINESS

For teams shipping to production.

149€

/mo + PAYG

Annual license, Business SLAs.
‍

Get started

Read the docs

Everything in starter, plus

99.5% monthly availability SLA

Business support, business hours (CET)

100 GB document storage included

Unlimited users — no per-seat fees

ENTERPRISE

For large organizations and the public sector.

Custom

Tailored deployment, Enterprise SLAs and white-glove support.

Everything in Business, plus

Dedicated environment: cloud, VPC, on-prem

SecNumCloud / regional options (MENA, APAC, US)

Managed deployment & setup

Custom data source connectors (1 included)

Dedicated CSM + Enterprise SLAs

Volume discounts

Optional custom AI engineering & integration

Need VPC, on-prem, or air-gapped deployment? See Enterprise deployment options

RAG was built for chatbots.
LightOn is built for agents.

Agents do not ask nicely. They dump raw PDFs, garbled tables, and off-domain queries into the same thread. The retrieval layer has to handle the input it gets, not the input you wish you had.

Retrieval

Hybrid retrieval

Dense, lexical, and late-interaction signals on one query. The index picks the right signal, not the developer. Built on LateOn and NextPlaid, our open-source ColBERT family.

Trust

Grounded by default

Every answer ships with the exact passage that supports it. Retrieval and reasoning are separable. Auditable by design, not retrofitted.

Infra

LLM-agnostic

Bring your own model. Open-source, commercial, or private. No lock-in on the inference layer. Your security policy dictates where inference happens.

Protocol

MCP-native

Drop LightOn into any agent that speaks Model Context Protocol. Single agent, multi-agent system, or business application integration. Same API.

Scope

Workspaces and ACLs at chunk level

Multi-agent systems need scoped corpora. Each agent gets its own workspace, its own collections, its own permissions. Access control is enforced at the chunk, not at the document. An agent never sees a single token it should not see.

What will you build?

Three endpoints. Hundreds of products waiting to be shipped.
Here are the ones builders are shipping right now on Console.

For Agent Builders

Ground your AI agent in your docs

Connect any agent (LangChain, LlamaIndex, custom) to LightOn via MCP. The agent calls /search , gets cited passages, generates grounded answers. No hallucinations, no orphan claims.

LightOn API + MCP + your LLM

See the MCP guide

For Product Teams

Add semantic search to your app

Replace keyword search in your SaaS with multi-vector retrieval. Upload docs via /parse, query via /search, surface results with citations. Two endpoints, no vector DB to manage.

LightOn API + your frontend

See the search recipe

For ISVs

Build a doc Q&A bot in an afternoon

Parse a folder of PDFs with /parse, index them in a workspace, query with /search, render the cited answer in your UI. Production-ready chatbot in <100 lines of code.

LightOn API + Next.js or Streamlit

See the Q&A Starter

BUILT ON OPEN RESEARCH

Our retrieval models are in your dependency tree.

Empower developers with a production-ready Multimodal RAG API running on your infrastructure. Integrate secure reasoning into your apps (CRM, ERP) without managing the complex AI stack.

50M

HuggingFace Downloads

916K

PYPI installs per mounth

2,345

GitHub Stars

3,845

HuggingFace Likes

Open-source models in production

LateOn

NextPlaid

PyLate

DenseOn

LightOnOCR-2

+8 more on HF →

Production-grade by default

The certifications and controls your security team will ask about. No checklist anxiety, no surprises in procurement.

Need deeper guarantees? See Enterprise security architecture → for VPC, on-prem, SSO, SCIM, and air-gapped deployments.

SOC 2 Type 1 certified

Audited annually

GDPR compliant

EU data residency by default

Audit logs & request tracing

Built in, queryable

Workspace isolation

Chunk-level ACLs enforced

Your data is never used to train models. Ever.

No exceptions, no opt-out clauses buried in the ToS.

Developer Experience

Built by Developers, for Developers

No enterprise procurement process to read your docs. No SDK rewrite every quarter. Just stable endpoints, real protocols, and a documentation that actually matches the API.

MCP-native

Connect any AI agent that speaks Model Context Protocol to LightOn in two lines. Single agent, multi-agent system, or app integration. The API is the same.

See the MCP guide

Comprehensive documentation

Full OpenAPI V3 specs, quickstart guides, copy-paste recipes for the most common workflows. Updated with every release.

Browse the docs

Standard protocoles

RESTful architecture. Consumable from Python, Node.js, Java, Go, or any HTTP client. No proprietary SDK required.

See the API reference

Don’t just take our word for it

Hear from some of our amazing customers who are building faster

The expertise of their tech team and the rapid evolution of the product, such as the hybrid search feature, put them at the forefront of innovation.

Jerome Lacaille - Emeritus Expert in Algorithms

Jérôme Lacaille
‍Emeritus Expert in Algorithms

Babbar needed an efficient SEO strategy enhancement through LLM technology to stay competitive in the dynamic SEO industry.

Sylvain Peyronnet - Currently working on IBOU (but still @Babbar)

Sylvain Peyronnet
‍Co-founder & search engine specialist

LightOn responded very quickly with tools that perfectly matched our needs, enhancing our document base and onboarding users without experience.

Achille Lerpinière
‍Chief Information & Technology Officer

The retrieval API for the agentic AI era

Parse a document. Extract any field. Retrieve with citations.
The same API key on Console, the same SDK on Enterprise.

LightOnOCR-2.
State-of-the-art parsing.

Define the schema.
Get JSON back.

Grounded retrieval
with citations.

Every endpoint is testable directly from Console.
Drop a file, get the response, copy the code into your project.

No seat licenses. No commitment. The free tier covers most prototypes.
Per-page pricing kicks in only when you scale.

Agents do not ask nicely. They dump raw PDFs, garbled tables, and off-domain queries into the same thread. The retrieval layer has to handle the input it gets, not the input you wish you had.

Hybrid retrieval

Grounded by default

LLM-agnostic

MCP-native

Workspaces and ACLs at chunk level

Three endpoints. Hundreds of products waiting to be shipped.
Here are the ones builders are shipping right now on Console.

BUILT ON OPEN RESEARCH

Production-grade by default

Developer Experience

No enterprise procurement process to read your docs. No SDK rewrite every quarter. Just stable endpoints, real protocols, and a documentation that actually matches the API.

MCP-native

Comprehensive documentation

Standard protocoles

Don’t just take our word for it

Start Building
Production-Grade AI

Parse a document. Extract any field. Retrieve with citations. The same API key on Console, the same SDK on Enterprise.

LightOnOCR-2.State-of-the-art parsing.

Define the schema. Get JSON back.

Grounded retrievalwith citations.

Every endpoint is testable directly from Console. Drop a file, get the response, copy the code into your project.

No seat licenses. No commitment. The free tier covers most prototypes. Per-page pricing kicks in only when you scale.

Agents do not ask nicely. They dump raw PDFs, garbled tables, and off-domain queries into the same thread. The retrieval layer has to handle the input it gets, not the input you wish you had.

Hybrid retrieval

Grounded by default

LLM-agnostic

MCP-native

Workspaces and ACLs at chunk level

Three endpoints. Hundreds of products waiting to be shipped. Here are the ones builders are shipping right now on Console.

BUILT ON OPEN RESEARCH

Production-grade by default

Developer Experience

No enterprise procurement process to read your docs. No SDK rewrite every quarter. Just stable endpoints, real protocols, and a documentation that actually matches the API.

MCP-native

Comprehensive documentation

Standard protocoles

Don’t just take our word for it

Start Building Production-Grade AI

Parse a document. Extract any field. Retrieve with citations.
The same API key on Console, the same SDK on Enterprise.

LightOnOCR-2.
State-of-the-art parsing.

Define the schema.
Get JSON back.

Grounded retrieval
with citations.

Every endpoint is testable directly from Console.
Drop a file, get the response, copy the code into your project.

No seat licenses. No commitment. The free tier covers most prototypes.
Per-page pricing kicks in only when you scale.

Three endpoints. Hundreds of products waiting to be shipped.
Here are the ones builders are shipping right now on Console.

Start Building
Production-Grade AI