Search papers, labs, and topics across Lattice.
This paper introduces ProRetrieval, a novel approach that transforms language models into retrieval orchestrators capable of synthesizing executable programs that combine SQL operations with vector-retrieval methods for text and images. By employing a hybrid domain-specific language (DSL), ProRetrieval effectively integrates structured constraints and semantic intents, overcoming limitations of existing hybrid retrieval systems. The results show that the 4B model significantly outperforms GPT-5.5 and Claude Opus 4.7 on new benchmarks, achieving higher accuracy in retrieving relevant information from complex datasets like Amazon products and Enron emails.
ProRetrieval outperforms state-of-the-art models by synthesizing executable retrieval programs that seamlessly combine structured and unstructured data queries.
Real-world retrieval often composes structured constraints with semantic intents over text and images through arbitrary Boolean logic. Existing hybrid pipelines such as reciprocal rank fusion or self-querying retrievers admit only a fixed form of composition, while recent reinforcement-learning retrievers train the language model as a query generator for a single backend, leaving the orchestration of heterogeneous retrieval paths outside its action space. We propose ProRetrieval, which recasts the language model as a retrieval orchestrator: given a natural-language query, it synthesizes an executable program in a hybrid DSL interleaving SQL operators over structured fields with vector-retrieval primitives over text and images, with SQL itself providing the logical algebra that fuses heterogeneous candidate sets. We train Qwen3-4B with GRPO and DAPO under a hierarchical four-term reward, and evaluate on two new benchmarks built from Amazon products and Enron email. Our 4B model surpasses GPT-5.5 (Hit@1 0.81 vs. 0.69 on e-commerce; 0.91 vs. 0.86 on email) and Claude Opus 4.7 and a comprehensive suite of retrieval, LLM-augmented, structured-query, and graph-based baselines. Code: https://anonymous.4open.science/r/ProRetrieval/; data: https://huggingface.co/datasets/anonymous-7219/ProRetrieval.
