Retrieval-Augmented Generation
AdoptTechniques
A method that combines retrieval with generation to answer using external context.
Why it's here
Placed in Adopt: 15 article(s) of evidence from 4 source(s), led by research-stage coverage, with 6 in the last 30 days. Confidence 81%.
Evidence (15)
- 7The New Stack·7/30/2026product_launchOpenAI and Elastic deepen enterprise AI integration
OpenAI and Elastic announced an expanded partnership that combines OpenAI reasoning models with Elasticsearch search, retrieval, and access-control capabilities. The integration targets enterprise AI context, observability, and security, with Elastic citing benchmark gains in retrieval efficiency, token reduction, and answer accuracy.
- 6The New Stack·7/23/2026researchPrompt Caching and Batched Pipelines to Cut RAG Costs
The article argues that naive retrieval-augmented generation (RAG) architectures often fail in production because synchronous ingestion, timeout limits, and rate limits do not scale to enterprise workloads. It recommends asynchronous uploads, micro-batching, and controlled embedding workers, and discusses prompt caching as one way to reduce costs without hurting accuracy.
- 6The New Stack·7/18/2026researchAI Agents Need Better Context Layers, Not Just Smarter Models
The article argues that many AI agent reliability problems come from weak infrastructure around the model, not from the model itself. It highlights Karpathy’s workflow as an example of compiling raw knowledge into structured, queryable context, then using retrieval and tools to help agents operate more reliably.
- 5The New Stack·7/16/2026researchWhy Smarter AI Caching Can Slow Systems Down
The article explains how exact-match caching with Redis helped early AI workloads but struggled once traffic and language variation increased. It then describes semantic caching with vector databases as a promising alternative, while noting that it can introduce higher latency, false matches, and operational complexity in production.
- 6The New Stack·7/15/2026product_launchWhat breaks when AI notetaker memory scales
The article examines how Plaud’s AI notetaker hit a scaling problem not in transcription quality, but in serving stored transcripts back to users quickly. Its original split between MySQL for metadata and S3 for large audio and transcript files worked at launch, but created latency and consistency issues when the product needed to behave like a real-time memory system.
- 6The New Stack·7/14/2026product_launchValantor launches enterprise visual intelligence platform after EyeLevel acquisition
Valantor acquired EyeLevel to launch its Enterprise Visual Intelligence platform, combining document intelligence with operational capabilities for unstructured enterprise data. The platform targets visually complex files such as PDFs, PPTX, and DOCX, and supports deployments in private cloud, on-premises, sovereign, and air-gapped environments.
- 6The New Stack·7/10/2026researchRetrieval quality becomes the key challenge in AI agent design
The article argues that many AI agent failures are actually retrieval and context-building problems rather than generation problems. It highlights the need for better ranking, tracing, and evaluation in retrieval pipelines, especially when agents search across chat histories, documents, code, and other complex sources.
- 6The New Stack·7/9/2026researchAutonomous ingestion errors poison a vector database
A fintech team describes how an autonomous PDF ingestion pipeline inserted incorrect metadata into a vector database, causing a chatbot to cite outdated and wrong financial data. The post-mortem says the root cause was treating probabilistic LLM extraction as deterministic, and that a second LLM-based validation step failed due to mutual confirmation bias.
- 8The New Stack·7/8/2026model_releaseRefiant unveils 10-million-token Protea model with swarm optimization
Refiant has launched Protea, a model with a 10-million-token context window that uses swarm-style optimization and context management to process very large datasets in one pass. The company says the approach could improve inference efficiency, reduce hallucinations, and help replace RAG for enterprise codebases and long-form clinical or email archives.
- 6InfoQ·7/7/2026product_launchHubSpot Scales Semantic Search to 20 Billion Vectors
HubSpot says its semantic search system has grown from a proof of concept into an internal platform used by more than 38 teams and now stores over 20 billion vectors. The service supports agents, retrieval-augmented generation (RAG), and contact deduplication, with retrieval quality and latency becoming more critical as agent usage increases.
- 6The New Stack·7/2/2026framework_updateWhy CI/CD Gates Fail for LLM Pipelines
The article argues that traditional CI/CD gates are insufficient for production LLM systems because model behavior is probabilistic and can drift without triggering hard failures. It proposes release gates based on baseline evaluations, drift detection, shadow validation, and cost/latency guardrails to catch silent regressions before users see them.
- 4InfoQ·7/1/2026researchGraphRAG Presentation on Smarter Retrieval with Knowledge Graphs
Cassie Shum presents the architectural evolution of GraphRAG and argues that strong data foundations are essential for advanced AI workflows. The talk contrasts GraphRAG with traditional vector RAG, highlighting its strengths in global context, multi-hop reasoning, and provenance through semantically structured knowledge graphs.
- 5Martin Fowler·6/16/2026researchHow Bayer Built a Reliable Agentic AI System for Preclinical Research
This case study describes PRINCE, a cloud-hosted platform built by Bayer and Thoughtworks to improve access to preclinical safety study data. The system evolved from keyword search into an agentic retrieval-augmented generation and Text-to-SQL assistant that can answer complex questions and draft regulatory documents, with engineering focused on transparency, recovery, observability, and human oversight.
- 6Hugging Face Blog·6/1/2026model_releaseJetBrains introduces Mellum2, a 12B MoE model
JetBrains has released Mellum2, an open 12B-parameter mixture-of-experts model trained from scratch on natural language and code. It activates only 2.5B parameters per token, is designed for low-latency, high-throughput inference, and is released under the Apache 2.0 license. The company says it is suitable for routing, RAG, summarization, sub-agents, coding workflows, and private deployment.
- 7Hugging Face Blog·4/9/2026framework_updateSentence Transformers Adds Multimodal Embedding and Reranking
Sentence Transformers v5.4 adds support for encoding and comparing text, images, audio, and video through the same API. The update enables multimodal embedding and reranker workflows for use cases such as cross-modal search, visual document retrieval, and multimodal RAG.