This work presents a production retrieve-then-match cascade that spends computation in proportion to difficulty: retrieval surfaces plausible matches, a lightweight text cross-encoder auto-resolves the high-confidence majority, and an agentic multimodal vision-language model settles the ambiguous remainder.
Abstract
Product linking, the entity-resolution task of mapping merchant product records to canonical catalog products, consolidates fragmented listings so downstream search, recommendation, and advertising see one clean entry per product. At marketplace scale, billions of noisy, multi-category records must be resolved against tens of millions of canonical products, where scoring every candidate with a single model is either too weak for the hard cases or too costly for the easy ones. We present a production retrieve-then-match cascade that spends computation in proportion to difficulty: retrieval surfaces plausible matches, a lightweight text cross-encoder auto-resolves the high-confidence majority, and an agentic multimodal vision-language model settles the ambiguous remainder by inspecting product images and issuing web searches for evidence that is in neither record. The cross-encoder is distilled from millions of dual-VLM-consensus labels, retiring human annotation from the training set, and is calibrated to auto-accept links at a 98% precision bar validated against a smaller operator-certified audit. The agent is a self-hosted open-weight model that reaches a closed frontier VLM's precision at a four-point recall cost (88% versus 92%) for roughly one-seventh the per-pair cost, with no fine-tuning. Per-pair cost spans nearly five orders of magnitude from the cheap cross-encoder to the frontier VLM, so escalating only the hard tail to the agent raises end-to-end link coverage from the cheap stage's 68% to 77%.
A unified pipeline deployed at Walmart that addresses both signal quality and model evolution is presented, and a Warm-Start Distillation technique that transfers domain-specific expertise from the legacy model to the new backbone is introduced.
Zhen Yang, Juexin Lin, Hongwei Shang et al.· Annual International ACM SIG...· 1 citation
Apple Music serves listeners across 150+ storefronts in dozens of languages, with a catalog that grows by hundreds of thousands of new tracks daily. At this scale, search recall on misspelled, transliterated, and cross-lingual queries becomes a dominant driver of session quality, particularly for tail queries that account for the majority of unique queries. We present a multilingual semantic retrieval system built on a 305M-parameter Siamese bi-encoder fine-tuned from GTE-multilingual-base with curriculum-scheduled multi-objective training. The model is integrated into the search stack via a hybrid retrieval architecture that blends dense nearest-neighbor results with the existing token-based index using quantile distribution matching, enabling deployment without retraining downstream rankers. Offline, the model achieves a 69% relative improvement in Hit@10 over GTE-multilingual-base. In a worldwide online A/B test, the system delivers a 2.28% relative conversion-rate (CR) lift overall, an 86% reduction in the no-result rate, and gains across every storefront with no observed regressions. The improvement is concentrated where it is needed most: tail queries see a 7.93% relative CR lift, compared with 0.89% for mid-frequency queries and 0.14% for head queries -- evidence that semantic retrieval improves recall on hard queries without disturbing well-served popular ones. To our knowledge, this is one of the largest search-quality improvements deployed on the platform.
Vishalaksh Aggarwal, Kevin Sebastian, Vivek Kanojiya et al.· 0 citations
Despite-encoder vision-language models expose a similarity interface that enables zero-shot retrieval but fails compositional constraints, this work proposes factored inference, which separates evidence extraction from constraint execution, and introduces LCSE (Logic-Constrained Score Editing), a training-free method that executes constraints externally using concept scores from frozen encoders.
S. Alshehri, Zhantao Yang, Han Zhang et al.· 0 citations
Knowledge-based Visual Question Answering (KB-VQA) requires models to retrieve visual entities matching the query image from large-scale encyclopedic knowledge bases and answer related questions. Existing multimodal Retrieval Augmented Generation (mRAG) methods rely on global visual features to match candidate entities, yet when the knowledge base contains numerous visually similar entities, the retriever struggles to distinguish them, populating the candidate set with visually similar but factually mismatched distractors. Since subsequent processing steps such as noise filtering are also confined to this fixed candidate set, errors from failed retrieval inevitably propagate to the final answer. To address these challenges, we propose MMAgent-R$^2$, an agentic mRAG framework that integrates visual reranking and active rejection as its internal verification mechanism. Visual reranking directly compares query and candidate images, capturing discriminative details beyond textual descriptions to precisely identify the target entity among similar candidates; active rejection discards unreliable results and retrieves additional candidates when no confident match is found, moving beyond the fixed candidate pool. We design a composite reward function with step-level verification rewards and achieve joint optimization of external retrieval, internal verification, and answer generation via GRPO training. Experiments on InfoSeek, E-VQA, and MMhops demonstrate that \ours{} achieves state-of-the-art performance, with particularly notable advantages in challenging retrieval scenarios and complex multi-image multi-hop reasoning tasks.
Tao Zhang, Ziqi Zhang, Zongyang Ma et al.· 0 citations
Translating natural language into graph query languages (NL2GQL) enables non-expert users to access graph databases, but supervised parsers depend on large annotated corpora and costly retraining whenever the schema evolves. Few-shot in-context learning offers a training-free alternative, yet its effectiveness hinges on demonstration selection, and conventional retrieval based on sentence-level matching and a single similarity signal often overlooks beneficial exemplars. We propose MGDR-SDR, a training-free approach that applies constituency parsing to derive three complementary representations of the utterance: a shallow abstraction preserving fine-grained lexical and syntactic detail, a deep abstraction exposing the intent-level syntactic skeleton, and a noun phrase list capturing concrete graph elements. These representations drive a multi-granular retrieval procedure that aggregates evidence across dense and sparse channels, yielding broader and more precise demonstration coverage. We evaluate MGDR-SDR on SPARQL generation over KQA Pro and Cypher generation over ZOGRASCOPE. Without any training, it reaches 86.58% Hits@1 on KQA Pro, surpassing strong supervised parsers and setting a new state-of-the-art among few-shot methods, and it transfers across GQL formalisms, attaining the best iid accuracy and a balanced generalization profile that surpasses all open-source few-shot baselines on ZOGRASCOPE. These results show that syntax-derived, multi-granular retrieval is an effective and generalizable strategy for few-shot NL2GQL semantic parsing.
Tong Li, Zhipeng Du, Xinyu Hu et al.· Machine Learning and Knowled...· 0 citations
Cross-document knowledge retrieval requires the retrieval and synthesis of evidence scattered across multiple passages linked by implicit or explicit inter-document references. Existing methods either treat the corpus as a flat collection of independent passages, thereby forfeiting the links that connect evidence fragments, or reconstruct those links by generating synthetic entity triples, an expensive process that introduces hallucinated graph edges. To address this, we propose BRIDGE (Bridging Retrieval with Inter-Document Graph Expansion). This four-stage passage-retrieval pipeline exploits inter-document reference signals to recover multi-hop evidence chains without requiring the construction of graphs from expensive pretrained language model calls. To bridge the gap between flat retrieval and complex query resolution, the pipeline begins by fusing lexical term frequency and the dense embedding signals via Reciprocal Rank Fusion using an INT8-quantised vector index that compresses the embedding store by $4\times $ . It then applies a cross-encoder model as an initial precision gate to discover supporting passages that share a minimal lexical overlap with the query. And subsequently, the system traverses an inter-document reference graph anchored on cross-reference structures using Personalised PageRank. Finally, a cross-encoder reranking is performed on the merged candidate pool to ensure consistent scoring across both retrieved and graph-expanded passages. We evaluate BRIDGE on three established multi-hop QA benchmarks: HotpotQA, MuSiQue, and 2WikiMultiHopQA. Without any dataset-specific tuning, our approach achieves Recall@2 of 79.8, 41.5, and 65.7, respectively, matching or exceeding baselines including HippoRAG (59.0, 41.0, 71.5) and ColBERTv2 (64.7, 37.9, 59.2) on two of three benchmarks. Our ablation analysis confirms that reference-graph expansion discovers an average of 7.0 additional relevant passages per query, while the final cross-encoder reranking prevents score displacement and maintains top-of-list precision.
