Néo Schobert

The GraphRAG engine

Not a plain RAG: an agent runs the investigation with a toolbox over the ontology, writes from the evidence alone, then checks its own answer before sending it.

1. 01

   Understand

   A follow-up becomes a standalone question, concept names are resolved in the open folder, and the agent asks for clarification when the question is ambiguous.

   ask_user

2. 02

   Plan

   In deep and research modes, a plan is written before acting and shown to the user.

3. 03

   Collect

   One agent drives the whole toolbox, calls several tools at once and relaunches the search when needed. In research mode, an orchestrator splits the question across four sub-agents in parallel.

   search_conceptsconcept_cardschema_propertysparql_querygraph_shortest_pathgraph_longest_chaingraph_top_degreealignment_correspondencesevidence_read

4. 04

   Write

   From the evidence alone: every fact cited by its IRI, in the language of the question, with a diagram of the subgraph used and its Turtle export.

5. 05

   Verify

   Numbers and IRIs cross-checked against the evidence, re-read by a judge model in advanced modes. Every diagram edge is looked up in the ontology: kept, flipped or dropped. An invented identifier triggers a targeted rewrite.

Architecture: one writer, stateless engines

Measured, not claimed

93 %

correct answers on a 42-question benchmark (60% with the original pipeline)

31 → 7 %

hallucinations, from the original pipeline to the agent

0.936

alignment F1 on OAEI Anatomy (best published: 0.941)

#1 / 5

alignment engines compared, on every test set

3,000+

automated tests across the six services