Optimizing Research Agents with Boolean Retrieval

Summary & Methodology Analysis

The Sieve architecture addresses the inefficiencies in standard deep-research agents that perform naive search-visit loops. By incorporating a Boolean Query Language (BQL) expression, the agent defines specific field constraints, such as targeting specific titles, sections, or metadata, before passing candidate documents to a ranker. This filtering step ensures that the subsequent ranking process, whether using lexical, dense, or hybrid methods, operates on a higher-quality candidate set. Once candidates are ranked, the system renders them into compact cards containing section headings and query-focused snippets of 25 tokens to maximize context density while minimizing throughput costs.

The system architecture is model-agnostic regarding the underlying retriever, supporting a wide range of configurations including Pyserini/Lucene BM25 for lexical retrieval alongside dense embedding models like those from the Qwen3 family and BAAI/bge-base-en-v1.5. This design allows developers to swap ranking components based on specific deployment needs. By restricting the agent to fetch only relevant sections instead of full documents, the approach significantly reduces the total token count per research task, effectively lowering the computational overhead and latency associated with long-context windows in deep-research workflows.

A primary limitation identified in the paper is the agent dependency on the language model to generate effective BQL expressions. Since the performance of the entire pipeline relies on the accuracy of the Boolean filter, suboptimal query formulation can lead to missed relevant documents or overly restrictive search results. The research evaluated this framework across datasets including BrowseComp-Plus, HotpotQA, and MuSiQue, confirming consistent accuracy improvements over traditional methods that lack the BQL candidate selection stage.