Google adds native multimodal RAG to the Gemini API's File Search
Google expanded Gemini API File Search with native multimodal RAG capabilities, enabling search across images and text with page-level citations. New research demonstrated that even frontier LLMs corrupt an average of 25% of document content during extended delegation workflows, while Anthropic signed a $1.8 billion computing deal with Akamai and Nvidia disclosed over $40 billion in AI equity commitments for 2026.
Developer Tools #
Gemini API File Search is now multimodal #
Google / Hacker News (100 points)
Google expanded its Gemini API File Search tool with native multimodal RAG support, powered by Gemini Embedding 2, enabling applications to process and retrieve across both images and text rather than text alone. The update adds custom metadata filtering for scoping queries to relevant data subsets and page-level citations that tie model responses to specific source pages. For developers building retrieval-augmented systems, multimodal search eliminates the preprocessing pipeline of converting images to text descriptions before indexing – a meaningful reduction in both complexity and information loss.
The Agent Development Lifecycle #
LangChain
Harrison Chase presents a systematic framework for building production agents: Build, Test, Deploy, Monitor, Iterate – emphasizing that the hard part is making this loop repeatable across teams rather than getting a single agent working. The post argues that shared evaluation frameworks, deployment infrastructure, and tracing systems are prerequisites for scaling from one agent to many, and that governance around costs, tool access, and asset discoverability enables rather than hinders production deployment. For teams transitioning from agent demos to sustained agent operations, this codifies the organizational infrastructure that separates prototype-stage efforts from production-grade deployments.
Research & Papers #
LLMs Corrupt Your Documents When You Delegate #
Philippe Laban, Tobias Schnabel, Jennifer Neville / Hacker News (417 points)
Using DELEGATE-52, a benchmark spanning 52 professional domains, the researchers tested whether 19 LLMs can reliably handle delegated document editing tasks and found that even frontier models (Gemini 3.1 Pro, Claude 4.6 Opus, GPT 5.4) corrupt an average of 25% of document content by the end of long workflows. Corruption worsens with larger documents and longer sessions, and tool use does not mitigate the problem. For anyone building agentic document processing pipelines, this quantifies a failure mode that practitioners have observed anecdotally – models are unreliable for autonomous multi-step document editing, and verification checkpoints throughout the workflow are not optional.
Aurora: A Leverage-Aware Optimizer for Rectangular Matrices #
Tilde Research / Lobsters
Aurora addresses a critical flaw in the Muon optimizer: “neuron death” in MLP layers, where neurons receive persistently small updates and fail to recover during training due to Muon’s inability to maintain uniform row-norm distributions on tall matrices. The solution combines orthogonality constraints with equal row-norm requirements through an alternating projection algorithm, achieving new state-of-the-art results on the nanoGPT speedrun benchmark and demonstrating 100x data efficiency gains when training a 1.1B-parameter model. For researchers and practitioners working on training efficiency, Aurora suggests that optimizer design at the matrix geometry level still has substantial low-hanging fruit.
OncoAgent: A Dual-Tier Multi-Agent Framework for Privacy-Preserving Oncology Clinical Decision Support #
Hugging Face Blog
OncoAgent is an open-source clinical decision support system that runs entirely on-premises on AMD Instinct MI300X hardware, using a dual-tier fine-tuned LLM architecture (9B and 27B models) with a four-stage Corrective RAG pipeline that retrieves from 70+ physician-grade oncology guidelines. The system enforces a zero-PHI policy – a dedicated redaction node strips protected health information before any text reaches an LLM – and achieves 56x faster synthetic data generation versus API-based approaches. For teams building AI systems in healthcare, the architecture demonstrates how to satisfy data sovereignty requirements while maintaining clinical-grade performance, though the results are from a hackathon project and warrant independent validation.
Funding & Business #
Nvidia has already committed $40B to equity AI deals this year #
CNBC / TechCrunch
Nvidia has invested over $40 billion in equity deals in 2026 alone, anchored by a $30 billion commitment to OpenAI alongside seven multi-billion-dollar investments in public companies including Corning ($3.2B) and IREN ($2.1B), plus approximately two dozen private AI startup rounds. Critics have flagged “circular deal” concerns – Nvidia investing in its own customers – but the strategy is producing outsized returns, with its $5 billion Intel bet now worth over $25 billion. The pace and scale of Nvidia’s investment activity is unprecedented for a chipmaker and positions the company as both infrastructure provider and financial stakeholder across the AI ecosystem, raising questions about competitive dynamics when your GPU supplier is also your investor.
Infrastructure #
Anthropic Signs $1.8 Billion Computing Deal with Akamai #
Bloomberg / Yahoo Finance / The Next Web
Anthropic signed a seven-year, $1.8 billion cloud computing agreement with Akamai – the largest deal in Akamai’s history – to address surging demand for Claude inference capacity after experiencing 80x growth in annualized revenue. The deal positions Akamai’s distributed edge network as the inference-side complement to Anthropic’s centralized training infrastructure on Google Cloud and SpaceX, addressing the increasingly fragmented challenge of serving low-latency inference across dozens of geographies. Akamai shares jumped 27% on the announcement, signaling that the market sees CDN-to-AI-cloud pivots as credible – and that inference distribution, not just training compute, is becoming the binding constraint for frontier AI deployment.
Threads to Watch #
Inference infrastructure is fragmenting beyond hyperscalers. Anthropic’s $1.8B Akamai deal follows a pattern where frontier labs are distributing inference across unconventional providers – SpaceX for connectivity, Akamai for edge – rather than concentrating entirely on AWS/GCP/Azure. Combined with Nvidia’s $40B investment strategy that includes data center operators like IREN, the inference layer is acquiring a more complex topology than the training layer ever had.
Document-level agent reliability remains an unsolved problem. The DELEGATE-52 finding that frontier models corrupt 25% of document content in extended workflows, combined with this week’s earlier research on agent scaffolding interference and overthinking/overacting failure modes, suggests that multi-step autonomous document processing is further from production readiness than single-turn capabilities would imply. Teams deploying document agents should budget for verification infrastructure proportional to workflow length.
Sources Unavailable Today #
These sources could not be fetched today. Links point to their homepages so you can check them directly.
- Stanford HAI – scrape: content_truncated