RAG & Retrieval

UNREAL proposes a model-native evidence selector built on a frozen language model, linking corpus retrieval with long-context inference. Its reported results show higher recall on large-scale Wikipedia retrieval and better accuracy after removing distractors from very long prompts.

CANOPY builds hierarchies over retrieved multimodal items and keeps evidence at different levels of detail according to the query. A critic can also request targeted retrieval when the current evidence is incomplete.

ScholarCatalyst moves scientific retrieval beyond topical similarity by asking which earlier papers could genuinely advance a research project. Its author-annotated results show that current agents still struggle to match expert judgment.

A new study explains when dense retrieval should use shared or separate query-document projections. Its bias-variance analysis shows that dual projections become more attractive as data grows or query and document geometries diverge.

D-RAC normalizes PDFs, Word files, presentations, spreadsheets, and scans into rendered pages before using one multimodal pass to produce retrieval-oriented Markdown. Chunk planning then operates on element identifiers instead of regenerating source text.

SELF-INDEX is a framework for continuously adapting search indexes without requiring people to diagnose every retrieval failure. It revises responsible index keys, validates changes, and simulates new queries to explore demands beyond observed workloads.

GLIE shows that the roughly thousand-vector representation used for each visual document page contains substantial geometric redundancy. A few normalized vectors can support search and on-demand reconstruction of the full embedding set.

OreoLook introduces three coordinated caches for conversational web search: recent session context, semantically equivalent queries, and URL embeddings. Its local search and caching stack runs on commodity CPU hardware, while final answer synthesis is handled by a remote inference provider.

Retrieval-augmented generation can fail not only when evidence is missing, but also when retrieved documents disagree. A new study decodes internal language-model signals to route inputs into answerable, insufficient, or conflicting cases before generation.

A study of spoken multi-hop question answering finds that richer retrieval structures can preserve higher absolute accuracy while becoming more sensitive to upstream speech-recognition mistakes. Corrupted query entities are the main path through which those errors spread.

Accel-backed Keenable has emerged from stealth with a $26 million seed round and a search index designed for AI systems rather than human readers. The startup says its index covers more than 100 billion documents and is already used by AI labs and inference providers.

Amazon DynamoDB now supports native vector search, allowing embeddings and application records to live in the same table. The feature simplifies AI architectures, but it does not make dedicated vector databases obsolete overnight.

A new study proposes IAR, a staged post-training framework for answering questions about a fixed document collection without retrieved passages at inference time. It separates knowledge injection, QA accessibility, and recovery of general capabilities.

Google is letting publishers embed a “Preferred Sources” button so readers can ask to see them more often across Search, Discover, and Google News. The feature offers a new distribution lever as AI-powered search reduces the need for users to visit source websites.

Poisoned retrieval contexts can make a language model appear more confident rather than more uncertain. A new study proposes D-SCAN, which detects attacks by tracking document-level attention collapse during generation.

Shopify says AI-driven discovery is sending more visitors and orders to its merchants, rather than cannibalizing traditional search. In Q2, AI-attributed traffic and orders to Shopify stores tripled year over year.

DeepVoyager-VL aims to move multimodal search beyond one-shot image understanding. It treats visual evidence as an active driver of intermediate reasoning, retrieval, and agent decisions.

UEmbed from Alibaba-NLP proposes a decoder-only multimodal embedding model that produces both sparse lexical signals and dense vectors in a single causal forward pass. The work targets search, RAG, and multimodal retrieval systems that currently rely on fragmented retrieval stacks.

Reddit CEO Steve Huffman is questioning whether Google’s AI Overviews create the same ecosystem value that traditional search links once did. The comments came as investors reacted sharply to uneven search referrals despite strong quarterly results.

A controlled scaling study of RAG paradigms finds that the best retrieval strategy changes as the corpus grows. While file-system-style agents lead on small corpora, BM25 overtakes them around the 10-million-token range and remains ahead at larger scales.

New Similarweb data cited by TechCrunch shows Google’s AI Overviews now appear in 43% of searches, up from 15% a year earlier. The shift suggests AI-generated answers are becoming central to how people find information online.