Trace AI agent decision rationale in a format regulators can understand
When you need to explain the reasons behind an AI decision to auditors or regulatory bodies, a simple 'similarity score' is not enough. This tool converts enterprise data into a structured graph, allowing you to trace which data led to which decision from the source. It does not replace LLMs or vector stores; instead, it adds a layer of meaning and causality beneath them.
Structured meaning instead of abstract similarity
Most AI agents retrieve information based on the similarity of embedding vectors, but this does not explain the source or causal relationship of the results. This system extracts entities and relationships from documents or databases to build a knowledge graph, attaching source information to each fact. Thanks to the graph structure, you can answer the question 'why did this result occur?' by following the data connections. Additionally, conflicting facts are automatically detected and handled without being arbitrarily overwritten.
Decision records and audit trails
Every AI decision is stored as a node in the graph, with explicit links between cause and effect. This allows you to search for past similar decision cases or trace back the impact a specific decision had later on. The generated records can be exported in the W3C PROV-O standard format, making them immediately usable for preparing audit materials in heavily regulated fields like finance or healthcare. The confidence score of the decision and the applied policy rules are also recorded to ensure transparency.
Integration with existing stacks
This tool operates by adding a layer on top of your existing LLMs or vector databases rather than removing them. It pulls data directly from major enterprise data platforms such as Databricks, Snowflake, and SAP, converting it into graph nodes. It supports various graph storage options, including Oxigraph and Neo4j, allowing you to switch backends without code changes. It also integrates with popular agent frameworks like Agno, CrewAI, and LangChain, making it easy to incorporate into existing workflows.
Considerations before adoption
This system focuses on tracing the flow of context provided externally to the model and the generated decisions, rather than explaining the internal reasoning process of the LLM. Therefore, it does not solve the black-box problem of the LLM itself. While graph-based reasoning and provenance tracking operate deterministically, specific benchmark figures for performance on large datasets or for certain industry standards are not specified in the README. Before adopting it, you should verify whether your data scale and regulatory requirements align with this structured approach.
