Agent-Based Optimization of Document Representations in Semantic Search

Semantic search is crucial in fields such as medicine, research, and data search, where adapting search systems to user needs is a challenge. This study proposes optimizing document representations using an agent-based approach that incorporates contextual embeddings of document titles, passages, and user queries. We evaluate eleven document representation strategies that combine embeddings of document title, document passages and relevant user queries using weighted linear combinations, averaging, and clustering. Each document is modeled as an agent, that uses aforementioned strategies to optimize its representation based on user interactions. The NFCorpus test collection simulates user feedback for validation, which is performed using nested cross-validation. We find that a linear combination of the title embedding, mean passage embedding and the mean over the known relevant queries offer the best trade-off between search-performance and index size. We further find, that incorporating embeddings of relevant user queries significantly improves the performance of representation strategies and the agent-based simulations. The agent-based system performs better in terms of nDCG@10 than the pure representation strategies but comes with a larger index size. The agent-based approach effectively adapts document representations, improving search performance over repeated query exposures. While some strategies increased index size, resulting in computational trade-offs, they generally provided better ranking performance. Future work may explore inter-agent collaboration and reinforcement learning to enhance agent logic.