Enterprise AI Analysis

Agent-OM consists of two Siamese agents: a Retrieval Agent and a Matching Agent, sharing a common memory. The Retrieval Agent extracts entities, metadata, and context, storing them in a hybrid database. The Matching Agent then finds correspondences, ranks, and refines results.

Key tools include Metadata Retriever, Syntactic & Lexical & Semantic Retrievers, Hybrid Database Store for retrieval, and Hybrid Database Search, Metadata Matcher, Syntactic & Lexical & Semantic Matchers, Matching Summariser, Matching Validator, and Matching Merger for matching.

Agent-OM was evaluated on three OAEI tracks: Conference, Anatomy, and MSE, comparing its performance against state-of-the-art OM systems. The system achieved results very close to the best long-standing performance on simple OM tasks.

Significantly, Agent-OM showed considerable improvements in matching performance on complex and few-shot OM tasks, demonstrating its generative power in diverse domains and levels of complexity.

Optimal similarity thresholds for Agent-OM were found to be in the range T∈ [0.90, 0.95], balancing precision and recall. For top@k values, k ∈ [3, 4, 5] was recommended to balance computational complexity and precision, ensuring enough candidates for the LLM to select from.

Higher similarity thresholds and lower top@k values can lead to higher precision but lower recall, while the inverse can lead to lower precision but higher recall, highlighting a crucial trade-off.

Current limitations include the high cost of API calls for commercial LLMs, potential LLM hallucinations (though mitigated by validation), and a trade-off between precision and recall that requires careful balancing.

Future work involves integrating multimodal input (e.g., ontology diagrams), extending to multilingual OM tasks (partially tested), and adapting to small language models (SLMs) for resource-constrained environments, while further improving planning complexity and self-correction.