LightOn

Overview

LightOn is a sovereign AI platform that enables organizations to deploy generative AI inside their own trusted environments while retaining full ownership of conversations, data, and workflows. The company provides three primary product lines: LightOn API, a production-ready multimodal RAG API; LightOn Enterprise, a secure AI chat solution for private knowledge bases; and LightOn Services, which delivers custom GenAI applications with professional support. With a GDPR-compliant architecture and support for on-premise and air-gapped deployment, LightOn is designed for European organizations and the public sector that require strict data control and sovereignty.

At the core of the platform is the LightOn API, which simplifies retrieval-augmented generation into three endpoints. The /parse endpoint uses LightOnOCR-2, a state-of-the-art open-weights OCR model, to convert scans, tables, handwriting, and multi-column layouts into structured Markdown in over 20 languages. The /extract endpoint accepts a user-defined JSON Schema and returns structured JSON for fields, entities, and key-value pairs, with asynchronous processing via webhooks and per-field citations. The /search endpoint delivers grounded retrieval by combining dense, sparse, and late-interaction signals in one query; the index dynamically selects the most appropriate signal, and every result is accompanied by the exact source passage. These components are built on the open-source ColBERT family—LateOn and NextPlaid—and are complemented by additional research models that have accumulated more than 50 million HuggingFace downloads and over 900,000 monthly PyPI installs.

LightOn’s architecture is explicitly built for agents rather than chatbots alone. Agent workloads often submit raw PDFs, garbled tables, and off-domain queries, so the retrieval layer is engineered to handle whatever input it receives. Trust is enforced by default: every answer ships with precise, clickable citations, and reasoning is kept separable from retrieval for auditability. The platform is LLM-agnostic, allowing organizations to bring open-source, commercial, or private models with no lock-in on the inference layer. It is also MCP-native, meaning it can be dropped into any agent that speaks Model Context Protocol. Enterprise features include conversational search and Q&A, massive multimodal RAG across millions of files, agentic reasoning chains for multi-step tasks, custom specialized agents per team, and universal data synchronization from SharePoint, Drive, Confluence, and file servers.

LightOn is deployed across demanding enterprise scenarios. Use cases include counterparty risk analysis across subsidiaries and languages, multi-jurisdiction contract review where exclusions and annexes may reverse protective clauses, and regulatory perimeter mapping from audits, inventories, and architecture diagrams. Strict governance and ACLs are enforced at the chunk level, mirroring native permissions from source systems and ensuring that an agent never sees a token it should not see. By combining sovereign deployment, grounded retrieval, and open research, LightOn empowers organizations to make faster, data-driven decisions without risking data leaks, and provides a European, GDPR-compliant alternative to US-based AI platforms.

Core capabilities

LightOn API

Production-ready multimodal RAG API with three unified endpoints—/parse, /extract, and /search—accessed via a single API key, with SDK parity between Console and Enterprise.

LightOnOCR-2

State-of-the-art parsing model that converts scans, tables, handwriting, and multi-column layouts into structured Markdown; supports 20+ languages natively and has open weights on HuggingFace.

/extract

JSON Schema-driven field extraction that returns structured JSON for invoice numbers, lease dates, claim IDs, contract clauses, and more, with citations per field and async webhooks.

/search

Grounded hybrid retrieval that combines dense, sparse, and late-interaction signals; every result includes the source passage that produced it, with chunk-level citations.

Sovereign deployment

Deploy GenAI in your trusted environment, including on-premise and air-gapped configurations, with complete data ownership and GDPR-compliant architecture.

LLM-agnostic inference

Bring your own model—open-source, commercial, or private—with no lock-in on the inference layer and security policy determining where inference runs.

MCP-native integration

Drop LightOn into any agent that speaks Model Context Protocol, from single agents to multi-agent systems and business applications.

Governance & ACLs

Workspaces and granular permissions enforced at the chunk level, mirroring native ACLs from source systems to guarantee data segregation and compliance.

Universal data sync

Seamlessly index and synchronize external knowledge sources such as SharePoint, Drive, Confluence, and file servers to keep the knowledge base continuously up to date.

Open research ecosystem

Retrieval models like LateOn, NextPlaid, PyLate, DenseOn, and LightOnOCR-2 are open source, with 50M+ HuggingFace downloads and 900K+ monthly PyPI installs.

Use cases

Counterparty risk analysis across subsidiaries and languages with clause-level citations

Multi-jurisdiction contract review to identify exclusions, annexes, and jurisdictional risks

Regulatory perimeter mapping from audits, inventories, and architecture diagrams

Internal knowledge management with secure AI chat over company documents

Massive multimodal RAG over millions of files, images, tables, and handwriting

Agentic workflows that require multi-step reasoning and tool chaining