Enterprise Knowledge Graph
Agentic pipelines that turn scattered public data on European enterprises into a graph you can query.
Finding data about European companies was never the hard part. The hard part was that it arrived in a dozen shapes from a dozen sources, and no two of them agreed on what counted as the same company.
So the extraction layer is agentic instead of a fixed pipeline. LLM agents with tool access decide what to pull from a source and how to normalise it, which is worth the complexity because every source has its own idea of structure. NLP models handle entity resolution. Deeper text analysis takes the classification passes that are too ambiguous to write rules for. Everything lands in a graph store (JanusGraph first, later AWS Neptune) with vector indexes alongside it, so the same data can be reached by traversal or by semantic search.
Most of the engineering effort went into failure modes rather than capability. An agent that can decide things can also decide them wrong, and a wrong entity merge is expensive to unpick later. So the work was in narrowing what an agent was allowed to conclude and making every conclusion checkable after the fact.
innoscripta SE · 2022–2024