Search “RAG”

Kakao Open-Sources Two Korean-Optimized AI Search Embedding Models on Hugging Face

On October 29, Kakao open-sourced two embedding models on Hugging Face, Kanana-2-3b-embedding (3 billion parameters) and Kanana-2-1.3b-embedding (1.3 billion parameters), targeting personalized AI agents and RAG scenarios. Kakao said that in internal evaluations, the two models outperformed llama-nemotron-embed-1b-v2 and Qwen3-Embedding-4B of comparable scale in Korean retrieval performance, with comparable English performance and higher inference efficiency, while memory usage and inference speed improved by more than tenfold compared with the previous-generation Kanana embedding models. The models use MRL to let developers adjust embedding vector size, and are released under the Kanana Open License, which permits commercial use.

Dynamic Island elements found in macOS 27 code, touchscreen Mac launch imminent

Tipster pdfu found 5 Dynamic Island visual assets in the newly added MenuBarAgent in macOS 27; some system compact views are already provided with the system, which may confirm that the next-generation MacBook Pro will replace the notch with Dynamic Island. Siri AI is expected to expand from Dynamic Island in a droplet-shaped design, and Bloomberg says the new models are priced $400 higher than existing products.

Mastra Launches Span Queries: Filter and Group by Span to Pinpoint Common Coding Agent Failures

Mastra has launched Span Queries, which can filter and group individual spans across all traces by conditions such as time range, type, name, status, error, and duration; it is available in Mastra platform, Studio, CLI, Client SDK, and observability endpoints. Each query can combine up to 100 conditions, return up to 1000 spans per page, group by up to 2 dimensions, and cover a time range of up to 365 days; the CLI requires mastra@1.36.0 or later, and the SDK requires @mastra/client-js@1.57.0 or later. Self-hosted projects can also use it when storage.observability uses DuckDB, PostgreSQL, or ClickHouse.

Google and XREAL Launch the First Android XR Wired Glasses, XREAL AURA

Google and XREAL have launched the XREAL AURA, the first wired XR glasses powered by Android XR. Orders open today in the United States, Canada, and South Korea, with sales in Japan starting November 2, at a starting price of $1,099. The glasses weigh less than 95 grams, use transparent lenses and a 70-degree field of view to project apps into the user's field of view, and offload heavy computing to a portable computing puck connected by cable; the puck is powered by the debut Qualcomm Snapdragon Reality Elite platform. Users can open multiple apps at once and pair them with a Bluetooth keyboard and mouse, or connect a laptop screen as a window via USB-C wired mirroring, and Gemini can provide assistance based on the content in the field of view and on the screen.

Google-Scale Software Engineering and CS Education: A Conversation with Titus Winters

On The Pragmatic Engineer podcast, Titus Winters looked back at Google's 2010 large-scale refactoring of 250M+ lines of C++ code, which was maintained by 10,000 engineers at the time; the team completed it in seven steps—parsing, matching, code generation, coexistence of old and new APIs, sharding, banning old patterns, and deleting old APIs—all well before the LLM era. He also discussed how Google, because Search could not be released monthly, mandated that every change come with tests, and how he took part in updating the ACM/IEEE undergraduate computer science curriculum guidelines, believing that almost all existing curricula should be torn down and rebuilt.

How to use retrieval logs to prioritize data quality issues

When enterprises import documents into RAG systems, they often lack relevance assessment just as they did when building Data Lakes back then, while the RAG system itself already records which data is actually invoked, which can be used to determine the priority of data maintenance. Bitkom data shows that 41% of enterprises with more than 20 employees have already put AI into production use, and another 48% are planning or discussing it, with most starting from internal document Q&A assistants. Tools such as Amazon Bedrock Knowledge Bases connect to Confluence, SharePoint, and entire Web domains through ready-made connectors and automatically complete chunking and indexing, but they do not pre-screen content; the more unorganized content in the index, the worse the retrieval hits, and the more teams will increase the k value, driving up the context cost passed to the large model with each request.

How can a transportation management TMS system use Mapbox to implement AI route planning visualization: multi-plan comparison and drag adjustment

TMS systems' AI route planning visualization uses the Mapbox Directions API's alternatives=true to obtain up to 2 alternative routes and compares them side by side with differentiated styles. Dragging waypoints triggers recalculation only on dragend, combined with 500ms debouncing to avoid consecutive requests; after changes, incremental metrics and the warnings field are used to indicate restricted roads and time window conflicts. The Mapbox Directions API supports at most 25 coordinate points per request, and under the driving-traffic profile at most 3 waypoints; for larger scales, the Optimization API or a self-hosted OSRM service must be used instead.

MemTensor Open-Sources Native Memory Model MemT-1.0, Available in 36B and 288B Tiers

The MemTensor team has open-sourced MemT-1.0, a model built on a native memory architecture, offering two size tiers—36B Nano and 288B Flash—and has released a 66-page technical report. The model connects a Native Memory Block in parallel alongside the standard Transformer Block, uses an erase gate, write gate, and forget gate to achieve token-level dynamic state updates, and extends memory capacity through a Mixture-of-Storage mechanism. According to the figures in the article, under a 256K history, VRAM drops from 5.06 GiB to 40 MiB, and first-token latency remains at 104–112 ms between 512 and 256K.

How IskayPet unified enterprise knowledge with Amazon Bedrock Managed Knowledge Bases

IskayPet built an enterprise knowledge assistant on Amazon Bedrock Managed Knowledge Bases, covering retail operations, veterinary care, HR onboarding, and procurement, supporting four languages across three countries without self-managing RAG infrastructure. The assistant ingests existing documents directly through the native SharePoint connector, combines hybrid search of semantic and BM25 keyword search, managed reranking, and document-level permission controls, and returns answers with source citations. The entire solution, from the store frontend to the AWS architecture deployment, was completed by Kiro with Claude Haiku 4.5.

Sarvam AI releases enterprise agent knowledge fabric architecture and benchmark results

Sarvam AI introduced a knowledge fabric built for enterprise agents, composed of two layers, a knowledge tree and a domain ontology, which compiles documents once into a structured layer that agents can navigate, query, and cite. On the Harvey LAB benchmark, answers using GLM 5.3 met 84.2% of expert grading criteria, higher than agentic RAG's 55.1%, with average task cost dropping from $0.64 to $0.28; on Sierra's tau-knowledge banking benchmark, it achieved accuracy comparable to the baseline with 64% fewer tokens. The knowledge tree is already used in the Sarvam cowork product, while the ontology is built domain by domain and enterprise customer by enterprise customer.

English translation is being prepared.

Microsoft unveils a range of new cloud sovereignty tools at the Paris Frontier Tour

At the Paris Frontier Tour, Microsoft CEO Satya Nadella and cloud business head Douglas Phillips unveiled a modular cloud sovereignty solution that lets enterprises decide for themselves the degree of control, transparency, and local execution for each computing task. A new Sovereign Control Panel aggregates Azure sovereignty signals, and a data guard and a Production Touch Ledger that records remote operations by Microsoft engineers are being rolled out gradually. Azure Managed HSM external key management is now generally available, and on the Azure Local side, DocumentDB, Blob Storage, GitHub Enterprise Local, and Foundry Local have been added. The German Informatics Society (GI) and others criticize such solutions as merely surface-level add-on controls, with the underlying software stack still controlled by US vendors.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Development Methods and Tech Stack for an AI English Writing Agent

An AI English writing agent is implemented through the collaboration of a multi-agent pipeline, and can be split into agents for outline and idea generation, grammar and spelling, discourse logic, style polishing, and evaluation and scoring, built on LangGraph or PydanticAI. The recommended tech stack is Python (FastAPI) paired with PydanticAI, LangGraph, and LlamaIndex; the main reasoning model can be DeepSeek-V3, Claude 3.5 Sonnet, or GPT-4o, with GPT-4o-mini and Qwen-Turbo for lightweight validation. The system also uses RAG to connect to authoritative corpora such as the Oxford and Longman dictionaries and the Academic Phrasebank, and sets up anti-hallucination and academic integrity guardrails.

English translation is being prepared.

Gergely Orosz on AI token usage being set as a KPI

The Pragmatic Engineer author Gergely Orosz responded to a plea for help from a middleware engineer in South Korea: their manager, influenced by the GenAI hype, set "AI token usage" as a KPI and repeatedly pressured the engineer because they rarely use AI. Orosz believes the manager is using a clumsy approach to push them to adapt to change, and says "excellent" and "refusing to learn new things" are not mutually exclusive—refusing to adapt is still unprofessional. He offers two paths: accept that AI is changing how work is done, or honestly bear the cost of refusal—the companies willing to hire you will keep shrinking.

English translation is being prepared.

HP OmniBook Ultra 16 hands-on first impressions: powered by Nvidia RTX Spark N1X

The HP OmniBook Ultra 16 is a 16-inch mobile workstation powered by an Nvidia RTX Spark N1X processor; the review unit features a 20-core CPU, a 6,144-core Blackwell GPU, and 32 GB of LPDDR5x-9400 memory. It has a 16-inch 3K (2880 x 1800) 120 Hz OLED touchscreen, runs Windows 11 26H1, and in testing delivered just over 6.5 hours of battery life; gaming performance beats Snapdragon X2, but fan noise is closer to that of an ordinary x86-64 PC.