RAG / 检索增强生成

Late-interaction retrieval (ColBERT, ColPali) scores a query against a document via the MaxSim operator. The standard PyTorch implementation materialises the full query-token x document-token similarity tensor only to reduce it away. At ColPali scale this is the single largest tensor in the pipeline (e.g. 21 GB in FP16 for 10K documents) and limits both candidate set size at inference and batch size during contrastive training. We present Flash-MaxSim (FM), an IO-aware fused GPU kernel that computes the same MaxSim scores without ever materialising the tensor, and extends the same principle to the training backward. At ColPali scale on A100 this cuts inference memory up to 9x and training memory by two orders of magnitude, unlocking candidate sets and contrastive batch sizes a single GPU could not previously reach. The kernel is a drop-in replacement, exact up to floating-point evaluation order under its stated FP32-accumulation protocol: rankings match the FP32 reference within 5e-4 of nDCG@10 on BEIR and REAL-MM-RAG. A separate INT8 path trades exactness for halved index storage at high fidelity. Released open-source.

We propose RCEM, a Robust Conversational search EMbedder that is additionally equipped with LLM's query reformulation capability without losing base model's generalization. Unlike prior conversational dense retrieval approaches that learn direct conversation-to-passage matching, RCEM aligns conversations, prepended by special token, to LLM-rewritten queries, while preserving the original embedding space. The unchanged embedding space automatically maps the rewritten-query to the relevant passages. As a result, RCEM (1) reduces overfitting by simplifying the alignment task from long passages to shorter rewritten queries, (2) eliminates the need for conversation-to-passage relevance labels for training, and (3) maintains its original embedding space that allows conversational queries against indexes built by original embedder without rebuilding them. Extensive experiments show that RCEM consistently outperforms prior approaches, achieving up to 30% improvement under distributional shift.

Community detection is a foundational capability in large-scale industrial graph analytics, powering applications such as fraud-ring discovery, recommendation systems, and hierarchical indexing for retrieval-augmented generation. Among modularity-based methods, the Leiden algorithm has been widely adopted in production because it delivers high-quality communities with connectivity guarantees. However, real-world graphs evolve continuously, and timely community updates are needed to keep downstream features and retrieval indices fresh. Meanwhile, existing dynamic Leiden approaches recompute the communities whenever their vertices and edges change, thereby almost degrading to near-full recomputation under frequent updates. To alleviate the efficiency issue, we study the efficient maintenance of Leiden communities in large dynamic graphs and present a novel algorithm, called Hierarchical Incremental Tree Leiden (HIT-Leiden). We first provide a boundedness analysis showing that prior incremental Leiden methods can incur essentially unbounded work even for small updates. Guided by this analysis, we propose HIT-Leiden which effectively reduces the range of affected vertices by maintaining connected components and hierarchical community structures. Extensive experiments on large real-world dynamic graphs demonstrate that HIT-Leiden achieves community quality comparable to the state-of-the-art competitors while delivering speedups of up to five orders of magnitude over existing solutions. The production deployment results show that HIT-Leiden meets stringent latency requirements under high-rate updates at scale.