Mistral ships Agentic Search: iterative retrieval beats one-shot RAG

Five file-style tools push SEC-filing QA to 86% on a third fewer tokens. Also: ChatGPT reads iMessage on your Mac, and Copilot drops six models Sept 1.

Nowline AUG 21 7:00 PM banner

Top AI stories from the last hour

Top AI stories from the last hour

Copy markdown

  • RAG becomes an agent that browses your documents

    Mistral's new Agentic Search hands the model five file-system-style tools — search, open, navigate, read, grep — so it hunts, opens, and verifies across a corpus instead of retrieving fixed chunks in one pass. On FinanceBench (368 SEC filings, 150 questions) accuracy jumped from 26.7% to 86%, and table-heavy OfficeQA Pro rose +45.6 points, while cutting tokens by up to a third and P90 latency by ~40%. No fine-tuning: it works over any model (tested on Mistral Medium 3.5 and GLM-5.2) via the Search Toolkit, Studio/Vibe, or a Search Starter App, cloud or on-prem.

  • ChatGPT can now read and send iMessages on your Mac

    OpenAI's Aug 20 desktop update adds an Apple Messages plugin, so ChatGPT on Apple Silicon Macs can read your iMessage, SMS, and RCS threads and draft or send replies. It's a concrete local-automation hook — summarize a thread, triage unread chats, or fire a reply from a prompt — shipping alongside site co-editing and editable Site URLs.

  • GitHub Copilot retires six models on Sept 1

    Copilot drops Claude Opus 4.5/4.6, Sonnet 4.5/4.6, Gemini 3.1 Pro, and Raptor Mini on September 1 — move pinned model choices to Opus 4.7/4.8/5, Sonnet 5, Gemini 3.6 Flash, or MAI-Code-1-Flash before then. One carve-out: annual individual subscribers keep Sonnet 4.6 so they retain a Sonnet option.

  • Elsewhere: Fish Audio text-to-speech is free on Vercel's AI Gateway

    Vercel put Fish Audio's TTS models on the AI Gateway free for 30 days (from Aug 19) — a no-cost window to wire voice into an app through the same Gateway you already route LLM calls through.