BM25
PulseAugur coverage of BM25 — every cluster mentioning BM25 across labs, papers, and developer communities, ranked by signal.
17 day(s) with sentiment data
-
Explicit document relationships boost LLM reasoning in RAG experiments
An experiment explored whether explicitly stated relationships between documents improve LLM reasoning in retrieval-augmented generation (RAG) systems when retrieval is held constant. The study found that providing expl…
-
On-device AI tool routing research highlights need for neural abstention
Researchers have explored on-device tool routing for AI assistants, distinguishing between tool selection and abstention (when no tool applies). Traditional methods using a single language model are expensive in terms o…
-
Quanta Python library unifies retrieval systems for RAG pipelines
Researchers have developed Quanta, an open-source Python library designed to simplify the integration of disparate retrieval systems into a single retrieval-augmented generation pipeline. Quanta unifies dense vector sea…
-
Lit3R system ranks 4th in literature-grounded QA task
Researchers from tus-nlp have developed Lit3R, a system designed for evidence-grounded question answering over scientific literature. This system integrates retrieval, reranking, and large language model components with…
-
VEKTOR enhances tool-calling and memory recall with new updates
VEKTOR has released updates across its platform, enhancing tool-calling capabilities for local LLM providers like Ollama, and extending this functionality to all send modes, including LIGHTNING, CASCADE, and COUNCIL/CRI…
-
New research reframes LLM prefix sharing as a sorting problem
A new paper proposes that optimizing the order of retrieved text pieces in retrieval-augmented generation (RAG) systems can significantly improve Large Language Model (LLM) serving efficiency. The research demonstrates …
-
ShopEase framework uses LLaMA 3.2 for enterprise customer support
A new research paper introduces ShopEase, a multi-agent framework designed for intelligent enterprise customer support. This system integrates six components, including intent recognition, CRM interaction, memory manage…
-
New framework unifies solar analytics, Q&A, and forecasting
A new research paper introduces Solar Intelligence, a hybrid framework designed to unify solar energy analytics, scientific question answering, and machine learning forecasting. This system integrates data from NASA POW…
-
New tools and research tackle GPU optimization for AI workloads
Several research papers and a new open-source tool address challenges in optimizing AI workloads on GPUs. COMPASS-ABS aims to reduce fragmentation in shared GPU clusters for deep learning training, improving resource ut…
-
LLM table descriptions worsen retrieval by increasing term frequency
An experiment with cataloging database tables using an LLM revealed that the generated descriptions, while accurate, worsened retrieval performance. The LLM described 1,245 tables, but this process led to a significant …
-
TF-IDF and BM25 identified as exact KL divergences in new arXiv paper
A new paper published on arXiv proposes that TF-IDF and BM25, two foundational methods for scoring query-document relevance in information retrieval, can be precisely interpreted as Kullback-Leibler divergences. This re…
-
Claude Developer Platform uses ranked search for tools, directory listing for memory
The Claude Developer Platform offers distinct approaches for managing overflowing tool catalogs and memory stores. For tools, the platform employs BM25 and regex-based search, returning a ranked shortlist of matching to…
-
HNSW vector search can silently miss data; tune ef_search
Vector search systems using Hierarchical Navigable Small World (HNSW) graphs can silently miss relevant data due to their approximate nature. The `ef_search` parameter, which controls the size of the candidate list duri…
-
Nautilus-Compass agent memory layer outperforms Mem0 on retrieval benchmarks
A new open-source memory layer for AI agents, named Nautilus-Compass, has demonstrated superior performance compared to Mem0 Agent Memory Framework on the LongMemEval-S retrieval benchmark. The Nautilus-Compass system a…
-
Hybrid Search Explained: Combining BM25 and Vector Embeddings for AI
This article explains the concept of hybrid search, which combines traditional keyword-based search (like BM25) with modern vector search. Vector search uses embeddings to represent text as numerical vectors, allowing f…
-
Nautilus-Compass launches LLM-free AI agent memory layer
Nautilus-Compass has released an open-source memory and reliability layer for AI agents that aims to improve long-term memory retrieval without relying on LLM extraction at write time. This approach stores raw text embe…
-
New Nepali passport QA dataset boosts retrieval performance
Researchers have developed a new question-answering dataset specifically for Nepali passport-related services, addressing the scarcity of resources for low-resource languages. The dataset was used to fine-tune transform…
-
New method audits object retention in grounded language models
Researchers have developed a new method called the Returned-Object Profile (ROP) to audit the effectiveness of grounded language-model pipelines. These pipelines involve selecting an object, retrieving relevant passages…
-
RAG techniques solve LLM context window limits for long documents
Large language models (LLMs) struggle with processing extremely long documents due to context window limitations, leading to slow responses, excessive costs, and inaccurate outputs. A practical solution involves impleme…
-
Developer replaces cloud LLM with local Ollama for cost savings
A developer has replaced the cloud-based generation component of their RAG chatbot with a local LLM, specifically Ollama running the qwen2.5-coder:32b model. This change was motivated by cost savings and privacy, tradin…