AI SEARCH · 2026-10-24
AI Search — Paper Daily, 2026-10-24
Today's ten items cover several stages of a question-answering pipeline. Three are about retrieval: on which queries hybrid retrieval actually does worse than dense retrieval alone; Lean Finder V2, which uses reinforcement learning to generate short queries for training Mathlib theorem retrieval; and SurveyGraph, which walks a graph of relations between methods to retrieve papers for each section of a survey. Four are about the RAG pipeline: how many chunk embeddings must be recomputed after a knowledge base is revised; CANOPY, which trims multimodal retrieved content down to the sentences, table rows or video clips a query needs; how to replace serial decoding with a single forward pass when an LLM picks sections level by level in hierarchical RAG; and GBOR-RAG, which extracts an evidence subgraph from a term–gene graph and hands it to an LLM to write the interpretation. On the answering side, JCA has the model check the question's conditions one by one so that it does not pick a distractor answer that simply appears more often in the retrieved results. Two industry items round out the list: Alibaba Cloud AI Search's full-stack optimisation from the retrieval engine to the embedding model, and Tencent Cloud TokenHub's interface that turns web search into a tool the model calls itself, with citations returned together with the answer.
4.85
Semantic Retrieval
Hybrid retrieval fuses BM25 and dense retrieval into one ranking, and most search engine and RAG framework docs default to reciprocal rank fusion (RRF). The authors keep a per-query ledger on 7 BEIR datasets: the best fusion method, score interpolation with weight 0.7 on the dense score, beats dense-only retrieval on mean nDCG@10 on every dataset, yet still ranks 8–21% of queries worse than dense-only. These losses concentrate on queries where only one retriever finds a relevant document; the other retriever then contributes only noise, and fusion pushes relevant documents that were already ranked high back down. Scaling the dense encoder from 22M to 596M parameters does not remove this loss. RRF beats dense-only only on datasets where the two retrievers are roughly equal, so the fusion method should be chosen from the measured gap between the two retrievers on the target corpus.
4.55
RAG Pipeline
RAG chunking schemes are usually compared on one static corpus, but knowledge bases keep changing. Each chunk is embedded on its own, so when an edit moves a chunk boundary, unchanged text must be re-embedded too. The authors measure this on 120 pairs of adjacent Chinese and English Wikipedia revisions, reusing an old vector only when a chunk's embedding input is byte-identical: non-overlapping fixed-size chunking recomputes 57–68% of the new version's embedding input, paragraph-first recursive chunking 38–56%, and content-defined chunking (CDC, which picks cut points from local text fingerprints) 43–54%, while evidence recall under a 2048-token retrieval budget differs by less than 2 percentage points among the three. When 1% of text is inserted at the start of an article, fixed-size chunking recomputes everything, because every window after the insertion point shifts, so no chunk matches the old version. For knowledge bases that change often, a chunker's recompute fraction should be reported alongside its retrieval quality.
4.25
Query Understanding & Rewriting
Mathlib is the math library of the Lean proof assistant. Queries from humans and coding agents searching it have median lengths of only 16 and 15 tokens, while the theorem paraphrases used to train earlier retrievers have a median of 82 tokens. Lean Finder V2 trains a query generator with reinforcement learning to write, for each theorem, short queries that the current retriever ranks poorly: no reward if the target theorem lands in the top 5, reward only for ranks 5 to 100, and the query must first be judged by an LLM to state the same fact as the theorem. The 600K+ generated short queries then train the retriever. On MathlibQR, 810 human-written short queries, Recall@10 rises from the previous best system's 0.780 to 0.893. Ablations show that rewarding difficulty without checking consistency makes retrieval worse, because a query may be hard simply because it drifted off target.
4.00
RAG Pipeline
In the text, tables, and videos that multimodal RAG retrieves, the part needed to answer is often just a sentence or two, one row, or one segment. CANOPY turns each retrieved item into a hierarchy (text down to sentences, tables down to rows, videos down to 30-second segments), scores every node against the query with an encoder fine-tuned on annotated evidence, and walks down from the root: if any child scores at least as high as its parent, it descends into those children and keeps comparing; otherwise it keeps the whole parent. So one item can keep several regions at different granularities, and pruning makes no LLM calls. When evidence is insufficient, a critic module writes a follow-up query and retrieves again. On 5 QA benchmarks over a 33M-item mixed corpus, with Qwen3-VL-8B as the reader, the five accuracy scores average 39.4, versus 32.9 for vanilla RAG. Ablations show the accuracy gain comes mainly from follow-up retrieval, while compression only saves input: compression cuts evidence tokens by 14.2–27.7% with no significant change in accuracy.
3.95
Multimodal Matching
Alibaba This is an article on the Alibaba Cloud developer community about full-stack optimization of AI Search, from the retrieval engine to the embedding model. The authors' starting point is that once agents go into production, search has to keep supplying the model with high-quality context, and the pipeline covers model understanding, vector and full-text recall, hybrid retrieval, fusion reranking and engine execution. The article says AI Search achieved breakthroughs on VDBBench vector retrieval, Big5 and Tantivy traditional retrieval, and MMEB multimodal document retrieval, backed by FalconSeek, an in-house Elasticsearch engine, and a multimodal embedding model. Its reference value is that it breaks an AI search pipeline into stages that can be optimized separately; the article summary gives no specific scores, so for a selection decision the raw results need to be checked on each leaderboard.
3.90
QA & Deep Research
Tencent This is the web-search API documentation for TokenHub, Tencent Cloud's model serving platform. With web search turned on, the model decides by itself whether to search and which keywords to search for while answering, and then generates the answer from the search results. There are two ways to integrate: the Responses API declares a web_search tool in the tools field and returns cited sources in structured annotations; the Chat Completions API turns search on with web_search_options and returns cited sources in search_results. It currently supports Hy3 and the DeepSeek V4 model family, search is priced in two tiers, lite and standard, the limit is 5 QPS, and it is only available in the Guangzhou region. For search teams, the reference value is the interface shape: search is packaged as a tool the model calls by itself, and cited sources come back together with the answer, so an in-house RAG service can align with these fields and make a later switch easier.
3.80
QA & Deep Research
CMU In retrieval-augmented QA, a model can pick the wrong answer even when the correct one is in the retrieved results: a competing candidate satisfies only part of the question's conditions (say, the question asks for the original Broadway actor and the candidate played the role in the film), but appears more often in the retrieved documents. Holding the question, candidates, and neutral evidence fixed and only raising the share of passages supporting the competing answer, the authors find that the rate of picking it rises in all 15 settings, by 1.4x to 9.7x. JCA (Joint Constraint Aggregation) first splits the question into atomic conditions, then uses the retrieved evidence to check which conditions each candidate meets, and finally compares all candidates' check reports together to choose the answer. With Llama4-Scout-17B, on the hard subsets where the competing answer is mentioned more often, JCA beats the strongest baseline, CoT, by 6.1–15.5 percentage points across three datasets. Given the same check results, comparing them side by side is more accurate than scoring by the number of conditions met, raising full-set accuracy by 3.8–9.3 points.
3.60
RAG Pipeline
In hierarchical RAG, at each level of a document's table-of-contents tree an LLM picks 3 of up to 20 sections to descend into. When this choice is executed with grammar-constrained decoding, every action token or syntax token costs one serial forward pass. The authors enumerate all selection prefixes reachable within 3 steps into a shallow tree: one prefill over the prompt fixes the first choice, then all possible continuation prefixes are concatenated into one sequence for a single forward pass, with a tree-causal mask so each prefix sees only the prompt and its own ancestors, and with the same position indices as serial decoding. Each prefix sees exactly the input it would see under serial decoding, so in exact arithmetic it gets the same probability distribution, and no token is generated at all. On Qwen2.5-7B and 3B, 84–89% of questions select exactly the same section sequence as serial decoding, and every mismatch is a near-tie under bf16 precision; selection latency drops by 35–47%.
3.55
Most automatic survey-writing systems retrieve papers as independent documents and cannot tell which paper builds on which method. SurveyGraph first builds a method-entity graph through four extraction layers (rules, citation contexts, shared entities, and LLM reasoning), with edges for 21 relation types such as "extends", "improves", and "applies to a new task", then retrieves papers for each survey section along one-hop graph neighbors. When the targets are core papers selected by the graph itself, graph retrieval reaches recall 0.373 versus 0.216 for flat lexical retrieval. With external targets drawn from papers that real surveys cite, at 5 papers per section, retrieval along typed edges alone reaches 0.153, below flat retrieval's 0.252 and below a TF-IDF similarity graph's 0.347. The first set of targets was itself chosen by this graph, so the near-doubling of recall cannot be taken directly as evidence of better retrieval, and the authors accordingly narrow their claim to retrieval results that are auditable and show how methods relate.
3.30
RAG Pipeline
Gene-expression studies commonly use over-representation analysis (ORA): testing whether a predefined gene set (a pathway, a biological process, and so on) contains more differentially expressed genes than chance would predict. Results are scattered across databases such as GO and KEGG, each giving its own ranked table, and readers must find the links between databases themselves. GBOR-RAG builds the 33,199 terms and 20,604 genes of 7 databases into one term–gene bipartite graph, runs the same hypergeometric test directly on the graph, and then uses prize-collecting subgraph extraction to find a compact evidence subgraph connected through shared genes, which an LLM uses to write the interpretation. Compared with conventional ORA, p-values have a Pearson correlation of 0.997, and median query time is 37x faster. Every term in the subgraph links to specific genes, so every generated claim can be checked against the evidence: on 42 GEO queries, 0.90 of claims are supported, versus 0.71 for GeneAgent.
EMNLP 2026