We model datasets as graphs (images, objects, labels, clusters) and pass only IDs + lightweight stats. Query-aware sampling and field-specific summaries mean the LLM usually sees ~2K tokens instead of 50M+, while tools still operate on full state. 🧵(4/5)

Our answer is a pass-by-reference architecture. Tools never stream raw metadata into the LLM. They return a tiny natural-language summary for the model + a rich artifact (IDs, filters, stats) that lives in system memory, not in the context window. 🧵 (3/5)

We recently launched the beta VL chatbot, which enables browsing visual dataset using natural language. Our newest blog post outlines our high-level architecture and efficient context management, for cases when your tools return more data than the context can hold. a 🧵(1/5):

See how Visual Layer is defining the future of visual intelligence and putting it into the hands of those who protect and defend. VL is in London with the @Mafat_IL delegation at DSEI 2025 today and tomorrow! Message us or stop by the booth to chat: lnkd.in/d8U8ZkMd

Visual Layer will be presenting our AI-native visual intelligence platform at DSEI UK this September, representing the Israeli defense tech scene under the auspices of @Mafat_IL

VL has been selected by @Mafat_IL, Israel’s Directorate of Defense Research & Development, as one of a few DefenseTech startups on display at DSEI UK 2025. Come see how VL helps defense organizations transform vast amounts of raw visual data into real-time operational insight.

Visual Layer will be at #MWC25 Barcelona next week, showcasing how we help organizations search, manage, and optimize massive visual datasets with AI-driven intelligence. 📍 Find us at: Hall 5 - Stand 5E61, 5G61

Visual Layer is now available on AWS Marketplace. You can now seamlessly integrate Visual Layer to explore, enrich, and extract insights from massive visual datasets all within the AWS ecosystem. Learn more here: aws.amazon.com/marketplace/p…

The new DeepSeek multimodal model, Janus-Pro-1B, is now available for inference on the Visual Layer platform. Try it out here: app.visual-layer.com/free-tr…

Thank you AI Day organizers Prof. Lior Wolf for inviting us to give a talk on how Visual Layer’s product is used in production to help structure index and search Iron Swords war footage. #AIDay