Search

[2611.07899] Persistent Retrieval for Conversational Agents: A Study of Retrieval Efficiency

The paper compares three retrieval designs for conversational agents and shows that a persistent retrieval subagent—one that maintains history and supplies evidence—reduces tokens and API cost and speeds up responses, at the trade-off of higher token usage versus non-persistent methods. It emphasizes maintaining retrieval state and enabling corrections from the subagent to improve answer generation.

What is Retrieval-Augmented Generation (RAG)? | Google Cloud

LLMs can hallucinate; grounding outputs with relevant facts from source materials via a long context window or RAG can improve accuracy and efficiency, reducing token usage and cost.

[2610.29722] Retrieval Augmented Generation Compositionality

A study using LASAGNA shows that combination effects in Retrieval Augmented Generation differ from isolated gains; performance depends on corpus size and the full system stack, urging evaluation of RAG techniques in complete pipelines.

[2611.03747] Enhancing Stock Market Information Retrieval with Retrieval-Augmented Generation

The document presents a Retrieval-Augmented Generation system for stock market information, showing that combining retrieval with generation improves accuracy and relevance, evaluated with ROUGE metrics on 100 queries and a latency of 1.16 seconds, suggesting the approach is promising for financial information retrieval.

[2610.26830] Competing Learning in Hybrid Architectures

Hybrid language models mix recurrence and attention; recurrence often learns retrieval faster, hindering attention. A staged training method is proposed to align learning, improving language modeling and retrieval-related tasks across contexts.

Retrieval’s bottleneck is moving from finding to selecting - kapa.ai

Retrieval workflows now face a shift from finding relevant chunks to selecting them, as models improve at search and query decomposition. In a two-model setup, finding gaps cause failures, and many gold chunks are never presented to the answer model, highlighting the growing importance of the selection step.

Google AI Mode: how to enable it and return to search

Translated title. Original: Modo IA de Google: cómo activarlo y volver a la búsqueda

Google AI Mode provides a mobile, conversational search experience using text, voice, and images, with AI-generated answers that summarize across sources. It can be accessed directly via google.com/ai or through the Google app, with optional experimental features and a Web filter as an alternative; users can switch back to traditional results, depending on device and account settings.

[2610.23226] Characterising and Analysing Retrieval-Augmented Generation Techniques

A conceptual framework is proposed to characterize and analyze Retrieval-Augmented Generation techniques, enabling pre-empirical filtering to select well-suited candidates before costly benchmarking. The framework emerges from qualitative literature analysis to define what RAG is and how effectively it works.

[2610.25096] Toward Agentic Retrieval-Augmented Generation for Enterprise Retail Workflows over Heterogeneous Data Sources

RetailBridge is a benchmark for testing agentic retrieval across diverse data sources in enterprise retail. RetailRAG provides a framework to evaluate LLMs on cross-source planning and grounding, showing multi-source chaining as the key bottleneck and underscoring the need for explicit cross-source orchestration.

Nahles sees these jobs as threatened by AI.

Translated title. Original: Diese Jobs sieht Nahles durch die KI bedroht

The text promotes frag noz, an AI search service that answers questions with summaries based on relevant articles from the noz.de archive. It offers a cancellable trial, followed by a one-year subscription costing €2 per week.