The use of Large Language Models (LLMs) and agentic systems applications increased exponentially in areas with operational and legal liability, such as health, financial, legal, and government.Concurrently, the AI systems’ architectures became more complex with the incremental addition of elements such as tools, memory, protocols, guardrails, and multi-step decisions to accomplish a diverse quantity of tasks, demanding the audit systems’ capability to also grow in complexity to perform verification of such systems. The analysis surpasses the use of system logs, providing the ability to reconstruct the AI system rationale in the steps taken, including data, tools, and theirrelationship with policies, controls, and regulatory obligations.This paper introduces VectaDB, an ontology-native metadatabase that unifies vector embeddings, graph structure, and typed ontologies to provide an audit-first substrate for AI applications, offering a robust analytical abstraction layer for AI auditing. VectaDB stores (i) the similarity search semantic representations, (ii) explicit relationships for provenance, causal, and lineage queries, and (iii) anontological layer for constraints and interpretability.VectaDB enables symbolic and semantic investigations due to the created representation of prompts, tool calls, retrieved evidence, artifacts, and outputs, linking the objects by provenance relations, retrieving previous similar incidents, tracing decision lineage, validating data quality constraints, and producing compliance evidence.
Roberto Williams Batista (Sun,) studied this question.