Lettria

Lettria is the enterprise knowledge graph platform of TURFU SAS, based in Paris, France. It unifies structured data and unstructured documents into ontology-driven graphs, giving the AI layer above it a traceable context layer. It is not sold to early tinkerers: GraphRAG is delivered project by project by embedded ontologists, while the underlying Perseus engine is opened up through its API, SDK and MCP.

Core capabilities

• Ontology building: automatically generates domain ontologies and governs their versions; the ontology is the 'contract' of the graph schema, and the official site claims it takes days.
• Text to graph: extracts entities, attributes and relations and maps them onto the ontology, with cross-source entity resolution and partial incremental refresh, so the whole graph does not have to be rebuilt.
• Graph retrieval: hybrid vector and graph retrieval returns entities, relations, paths and source citations, so answers stay traceable.
• Document parsing: parses PDF, Office files, spreadsheets, images and audio, rebuilds tables and lists, and outputs full text plus chunks.
• Portable export: graphs can be exported to Neo4j, FalkorDB, Neptune as well as TTL, Cypher and SPARQL, and ontologies support OWL and SKOS.

Where it fits Mid-size and large enterprises in regulated industries in Europe (especially France): banking and insurance compliance and ESG disclosure, literature structuring for pharma companies and hospitals, law firm contract libraries, industrial manuals; also teams that already run a graph database and want to engineer extraction quality. Individuals and small teams looking for sign-up-and-go should look elsewhere.

What sets it apart Competitors sell retrieval; Lettria sells ontology governance: embedded ontologists, quarterly ontology refresh, SSO/SAML, audit logs and SIEM export, sovereign cloud residency — the gap that open-source GraphRAG cannot fill. Its own Text2KGBench-LettrIA (19 ontologies / 14,882 triples) underpins the 99.95% output reliability and Avg F1 0.89 of a fine-tuned Mistral Small 3.2.

Editorial take The strategy is coherent and the evidence solid, but it is hard to evaluate: all three GraphRAG tiers are custom-quoted, and the only public number, the PCU unit price, is undefined site-wide, so buyers cannot budget on their own; the performance claims contradict the company's own leaderboard, and third-party word of mouth is close to nil, with the 13 five-star reviews on G2 dating from 2023 and an older product. A serious candidate for an enterprise shortlist, but not something to try today.