Make classic BM25 search smarter without rebuilding expensive neural indexes — by optimizing query rewriting one token at a time. A genuinely clever approach 🔎 Title: STORM: Stepwise Token Optimization with Reward-Guided Beam Search URL: 🔎 Overview STORM trains a query-rewriting model guided by retrieval quality. At each generated token, it scores candidate expansions against a BM25 index, concentrating exploration on the vocabulary that actually improves search. ❓ Challenges Solved Modern retrieval leans on dense and learned-sparse neural models, while lexical methods like BM25 are fast but weak on synonyms and paraphrases. ・Dense neural retrievers need expensive index rebuilds whenever the model changes ・LLM query rewriting tends to produce well-formed but retrieval-ineffective or harmful terms ・Training gives only delayed sequence-level feedback, obscuring which individual terms actually helped 💡 Methodology & Proposed Approach ・Self-supervised training via reward-guided beam search driven by retrieval performance ・At each token, candidate expansions are scored against BM25 and low performers pruned ・This turns retrieval metrics into token-level signals, focusing search on effective vocabulary ・Using BM25 indexes means no neural index rebuilding — infrastructure stays light 📊 Experimental Results ・0.6B-8B models match or exceed competitive LLM rewriters ・Maintains BM25's speed advantage ・The 8B variant rivals much larger proprietary systems ・Zero-shot transfer to 18 languages (MIRACL) beats dedicated multilingual dense retrievers on average 🌍 Use Cases It fits search stacks that want to avoid index-rebuild costs, and systems that need cheap multilingual boosts. Since it lifts performance while keeping an existing BM25 pipeline, it's an easy-to-adopt answer for teams running search in production. #Retrieval# #BM25#

Most of the attention on full-text search goes to BM25. Query a part number, get that part number back first. A query like "why does PROD-001 overheat under load" has a literal string and a question in it though. That's where text-match filtering comes in. It narrows results to the documents that contain the literal string, and semantic search ranks what's left, all from one index. Full-text search is GA in Pinecone Database, text-match filtering included. 🔗

If search-agent accuracy has plateaued, the next thing to cut is context tokens. Context as Infrastructure: Knowledge Compilation for Retrieval in Agentic Systems with Harnesses A harness using only BM25 and grep reportedly reaches 91% on FRAMES. What's left is wasted tokens from rereading the same documents. The paper proposes "knowledge compilation": adding context offline with an LLM. 🧾 Highlight 1: claim-evidence pairs Each unit is a normalized claim, a verbatim evidence span, and a source offset. Claims without a locatable span are not stored. When retrieval finds a chunk, the chunk is replaced by these pairs before it enters the context window. 📉 Highlight 2: structure didn't help, substitution did Summary trees, outlines, wikis and graphs showed no detectable accuracy gain. Claim-evidence substitution cut query tokens by 38% on average on FRAMES (N=200) with accuracy within noise. The cost is more retrieval rounds and higher latency, which a decision gate helps contain. 🧮 Highlight 3: a payback formula With gate acceptance rate p_h, break-even is about 2%. The measured average implies roughly 43%, so a compile step costing 10 queries pays back after about 26 queries. Caveats: one benchmark, single runs, no confidence intervals, and the false-accept audit hasn't been run. Treating context as infrastructure feels practical to me. #RAG# #AgenticAI#

AI agents need fast, accurate search at massive scale. Lakebase Search is now GA, bringing scalable vector and BM25 full-text search directly into Postgres. - New frontier for price-performance, latency, and recall on VectorDBBench - 4x cheaper than running pgvector for the same workload - True pay-per-use with zero compute cost when idle

I build small, practical AI systems. On-device chatbots without LLM APIs. Embeddings, BM25, small policy networks, explainable responses. 100+ free browser-side tools for AI, math, dev, security, SEO, and weird utility work. GitHub:

Is re-ranking the top 1000 really the best choice for every query? Optimal Re-Ranking Depth ❓ Does the best depth vary by query? 💡 Yes. The optimal re-ranking depth can be defined per query as the shallowest depth that reaches its best effectiveness, and for some queries deeper re-ranking actually hurts. On MS MARCO, an oracle choosing the depth improves DCG@10 by over 7% while cutting the average depth about 5x. ❓ Can we predict it in advance? 💡 Eleven existing query performance prediction (QPP) methods were mostly poor at it. A promising alternative is "IR-DCG@10": the DCG@10 of the first-stage ranking, computed from LLM relevance judgments. ❓ How much can we save? 💡 With BM25 and Mono-T5, average depth dropped from 511 to 331 (about 3x) at the same effectiveness as full re-ranking; the best QPP baseline gave about 2x. With SPLADE the optimal depth is already around 20, and IR-DCG@10 gives about a 15x reduction. ❓ Any caveats? 💡 Judgments come from a single LLM (GPT-4o-mini) over shallow top-10 pools, the authors call the results preliminary, and a gap to the oracle remains. For deployment, I'd want to know the cost of needing LLM judgments at prediction time. #InformationRetrieval# #Reranking#

TL;DR Teaching Qwen3 models (1.7B-8B) tool selection over 10,000+ tools with RL beats prompting the 402B Llama-4-Maverick on the same job. Title: Training Small Language Models for Transferable Tool Selection URL: Key points ・🧩 Plan-Retrieve-Select: split the instruction into substeps, retrieve candidates, pick one tool per substep ・🧹 Noisy ToolBench G3 gold labels are filtered with a teacher model down to 20,093 examples ・🎯 SFT then GRPO: rank-weighted F1 for Plan, exact match for Select ・📈 ToolBench G3 success: 41.5% at 8B vs. 30.6% for the best prompted Maverick ・🔁 Swap BM25 for the dense retriever COLT and accuracy improves with no retraining ・🧪 A learned selector adds about 13 points over a BM25 top-1 rule ・⚠️ Planning transfers poorly to TaskBench, where a large prompted model still wins A good example of why a small, purpose-trained model can beat a giant prompted one on cost, speed and accuracy. #LLMAgents# #ToolUse#

Great RAG paper from IBM. There are some really good ideas on how to solve common RAG issues. It's well known that retrievers chunk long documents by length, which discards the hierarchy the document already has. So they propose using a table of contents. A table of contents helps to encodes exactly the global structure that chunking throws away. STAIR uses that table of contents as the addressing scheme for a generative retriever, so the model stores and retrieves information from its own parameters against a structure the corpus supplies. On SearchTome, it reaches Recall@1 of 82.6 percent against 76.9 percent for a fine-tuned Differentiable Search Index, a statistically significant gap, with BM25 at 59.5 percent and DPR at 68.7 percent. Hallucination stays below 0.05 percent, which is the standing objection to generative retrieval and the reason grounding the address space in a real hierarchy is worth the extra structure. The ablations also show it generalizes where very few training samples exist. Paper:

Sentence Transformers v6 has been released, and it is centered around Multi-Vector Embedding Models, what does that mean? Usually, when you build a semantic search or RAG-based system, you've probably used the @OpenAI embedding API or similar, as it allows to quickly index and search through a set of documents. Such embedding models are called "bi-encoders", as they typically encode 2 things ("bi"): encode each document or chunk, encode the query, and then compute pairwise cosine similarity between each (document, query) pair to retrieve the top-k ones. This is known as vector search or semantic search, and typically achieves better results compared to traditional keyword search, which relies on the classic BM25 algorithm. In 2020, researchers at @Stanford came up with something better than bi-encoders, called ColBERT. The idea is to create many vectors for each document, and many vectors for each query. One creates a vector for each token (a word or part of a word) of the query, and a vector for each token of each document. Next, to find the best matches, one computes the so-called MaxSim similarity. This is illustrated in the animation below. Unlike traditional vector similarity metrics that operate on pairs of single vectors, MaxSim computes similarity between sequences of vectors. The key insight here: each query token finds its best match in the document, then we sum. This enables fine-grained semantic matching and avoids averaging. These "multi-vector" embedding models are also called "late-interaction" models, as they keep multiple vectors for each document and query, delaying the matching step until the very end. Over the last year, researchers at @LightOnIO trained some very impressive late interaction models, which they openly released on @huggingface. Today, the Sentence Transformers library, which is the go-to library for open-source embedding models, added first-class support for them, as well as for a variety of SOTA late interaction models trained by @mixedbreadai, @liquidai and more. This means that you can now also train or fine-tune your own late interaction models very easily. Various vector databases like @elastic and @qdrant_engine already include support for multi-vector embedding models. This integration will definitely boost adoption of multi-vector embedding models by the industry. Read more here: - ColBERT paper: - Blog: