ApertureData

New website, fresh documentation, and a MAJOR release on deck! Managing multimodal data is a challenge, but we’re making it easier with our newly launched ApertureDB Cloud. With experience deploying ApertureDB to Fortune 100 customers, we now bring the features to , ApertureDB Cloud which enables faster insights and streamlined workflows. Big thanks to @VentureBeat & @mr_bumss for featuring the release!! Key Figures: •35x faster at mobilizing multimodal datasets •2-4x faster than other some open-source vector search solutions •66% of enterprise data remains unused—time to change that. venturebeat.com/data-infrastru…

I was learning about 10+ new AI tools a day and realized I had no way to filter the noise. The spark for a better map came from the agentic stack sessions at @PWVentures . Watching founders build real-world "plumbing" forced a shift in perspective. 1/4 🧵

I’ll be at @NVIDIAGTC next week. Jensen Huang’s talked about AI as a “five‑layer cake”: energy → chips/compute → cloud infrastructure → models → applications, and it's becoming the industry’s shared map. What’s interesting is how quickly the software layers are exposing a gap: 𝘁𝗵𝗲𝘆 𝗻𝗲𝗲𝗱 𝗮 𝗱𝗮𝘁𝗮 + 𝗺𝗲𝗺𝗼𝗿𝘆 𝘀𝘂𝗯𝘀𝘁𝗿𝗮𝘁𝗲 𝘁𝗵𝗮𝘁 𝗴𝗶𝘃𝗲𝘀 𝗔𝗜 𝗰𝗼𝗻𝘁𝗶𝗻𝘂𝗶𝘁𝘆, 𝗴𝗿𝗼𝘂𝗻𝗱𝗶𝗻𝗴, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗮𝗯𝗶𝗹𝗶𝘁𝘆 𝘁𝗼 𝗼𝗽𝗲𝗿𝗮𝘁𝗲 𝗼𝘃𝗲𝗿 𝗿𝗲𝗮𝗹‐𝘄𝗼𝗿𝗹𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁. Verticalization is accelerating this shift. Physical AI is pushing it even faster. Stateless prompts don’t cut it when systems need to see, remember, and act across time. A few questions I’m bringing into GTC: • Do small, domain‑native models win once memory becomes the anchor, or do frontier models keep stretching upward? • How long can teams absorb token costs before architecture becomes the real constraint? • And what exactly is “memory” now — a log file, a vector store, or the core infrastructure layer that binds the stack? At @ApertureData, we’ve been thinking about this layer for a long time. 𝗜𝗳 𝘆𝗼𝘂’𝗿𝗲 𝗲𝘅𝗽𝗹𝗼𝗿𝗶𝗻𝗴 𝗶𝘁 𝘁𝗼𝗼, 𝗹𝗲𝘁’𝘀 𝗰𝗼𝗻𝗻𝗲𝗰𝘁.

The PoC days of AI are over. 𝗧𝗵𝗲 𝘁𝗶𝗺𝗲 𝗳𝗼𝗿 𝗽𝗿𝗼𝗱𝘂𝗰𝘁𝗶𝗼𝗻 𝗶𝘀 𝗻𝗼𝘄! Across industries, and especially inside AI-native tech companies, the same cracks are showing: •   Retrieval latency •   Multimodal sprawl •   Glue code overload •   Graph + vector fragmentation Before context graphs and AI memory, getting the data foundation layer right is the first order of business today. @ApertureData

Hey, AI models are everywhere, getting smarter as we speak, butttt we talk far less about how 𝘀𝗰𝗮𝗹𝗮𝗯𝗹𝗲, 𝗽𝗲𝗿𝘀𝗶𝘀𝘁𝗲𝗻𝘁, 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗺𝗲𝗺𝗼𝗿𝘆 𝗶𝘀 𝗱𝗲𝘀𝗶𝗴𝗻𝗲𝗱 𝗶𝗻 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀. In practice, “memory” shows up as a combination of things teams already struggle to manage: stored embeddings, structured metadata, relationships between entities, and the raw multimodal assets themselves. When these pieces live in different systems, memory becomes fragmented. Context is hard to retrieve consistently. Long-running workflows become brittle. Treating memory as infrastructure means designing the data layer so that embeddings, metadata, and multimodal content can be stored, queried, and related in one place. 𝗜𝘁 𝗺𝗲𝗮𝗻𝘀 𝘀𝘂𝗽𝗽𝗼𝗿𝘁𝗶𝗻𝗴 𝗿𝗲𝘁𝗿𝗶𝗲𝘃𝗮𝗹 𝗽𝗮𝘁𝘁𝗲𝗿𝗻𝘀 𝘁𝗵𝗮𝘁 𝗴𝗼 𝗯𝗲𝘆𝗼𝗻𝗱 𝘀𝗶𝗻𝗴𝗹𝗲 𝗾𝘂𝗲𝗿𝗶𝗲𝘀, 𝗮𝗻𝗱 𝗺𝗮𝗸𝗶𝗻𝗴 𝗶𝘁 𝗽𝗼𝘀𝘀𝗶𝗯𝗹𝗲 𝗳𝗼𝗿 𝗮𝗴𝗲𝗻𝘁𝘀 𝗮𝗻𝗱 𝗮𝗽𝗽𝗹𝗶𝗰𝗮𝘁𝗶𝗼𝗻𝘀 𝘁𝗼 𝗯𝘂𝗶𝗹𝗱 𝗰𝗼𝗻𝘁𝗲𝘅𝘁 𝗼𝘃𝗲𝗿 𝘁𝗶𝗺𝗲 𝘂𝘀𝗶𝗻𝗴 𝘁𝗵𝗲 𝘀𝗮𝗺𝗲 𝘂𝗻𝗱𝗲𝗿𝗹𝘆𝗶𝗻𝗴 𝘀𝘆𝘀𝘁𝗲𝗺. As AI systems move from one-off calls to multi-step, multi-agent workflows, 𝗺𝗲𝗺𝗼𝗿𝘆 𝘀𝘁𝗼𝗽𝘀 𝗯𝗲𝗶𝗻𝗴 𝗮 𝗯𝗼𝗹𝘁-𝗼𝗻. 𝗜𝘁 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗽𝗮𝗿𝘁 𝗼𝗳 𝗵𝗼𝘄 𝘁𝗵𝗲 𝘀𝘆𝘀𝘁𝗲𝗺 𝗶𝘀 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲𝗱 𝗳𝗿𝗼𝗺 𝗱𝗮𝘆 𝗼𝗻𝗲. And that’s where our inspiration and direction at @ApertureData is coming from, as we continue to design and refine the infrastructure for multimodal data, vector search, and relationship-aware queries in a single layer.

𝗖𝗼𝗻𝘁𝗲𝘅𝘁 𝗴𝗿𝗮𝗽𝗵𝘀 are getting a lot of attention, and for good reason (@JayaGup10 @FoundationCap )! They point to a future where human and agent decisions can be captured, understood, and revisited with far more fidelity than today’s systems allow. 𝘉𝘶𝘵 𝘵𝘶𝘳𝘯𝘪𝘯𝘨 𝘵𝘩𝘢𝘵 𝘷𝘪𝘴𝘪𝘰𝘯 𝘪𝘯𝘵𝘰 𝘳𝘦𝘢𝘭𝘪𝘵𝘺 𝘳𝘦𝘲𝘶𝘪𝘳𝘦𝘴 𝘮𝘰𝘳𝘦 𝘵𝘩𝘢𝘯 𝘢 𝘯𝘦𝘸 𝘢𝘣𝘴𝘵𝘳𝘢𝘤𝘵𝘪𝘰𝘯. If context graphs are going to function as a system of record for reasoning, 𝗼𝗿𝗴𝗮𝗻𝗶𝘇𝗮𝘁𝗶𝗼𝗻𝘀 𝗻𝗲𝗲𝗱 𝘁𝗵𝗲 𝘁𝗲𝗰𝗵𝗻𝗶𝗰𝗮𝗹 𝘀𝘂𝗯𝘀𝘁𝗿𝗮𝘁𝗲 𝗮𝗻𝗱 𝘁𝗵𝗲 𝗰𝘂𝗹𝘁𝘂𝗿𝗮𝗹 𝗽𝗿𝗮𝗰𝘁𝗶𝗰𝗲𝘀 𝘁𝗼 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝘁𝗵𝗲𝗺. ApertureDB provides the data layer. Agent frameworks provide the reasoning. Integrations provide the raw material. The harder part is 𝗰𝘂𝗹𝘁𝘂𝗿𝗮𝗹: 𝗹𝗲𝗮𝗿𝗻𝗶𝗻𝗴 𝘁𝗼 𝗷𝘂𝘀𝘁𝗶𝗳𝘆 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻𝘀, 𝗮𝗻𝗻𝗼𝘁𝗮𝘁𝗲 𝗿𝗲𝗮𝘀𝗼𝗻𝗶𝗻𝗴, 𝗲𝘃𝗮𝗹𝘂𝗮𝘁𝗲 𝗼𝘂𝘁𝗰𝗼𝗺𝗲𝘀, 𝗽𝗿𝗼𝘁𝗲𝗰𝘁 𝘀𝗲𝗻𝘀𝗶𝘁𝗶𝘃𝗲 𝗰𝗼𝗻𝘁𝗲𝘅𝘁, 𝗮𝗻𝗱 𝘁𝗿𝗲𝗮𝘁 𝗱𝗲𝗰𝗶𝘀𝗶𝗼𝗻 𝘁𝗿𝗮𝗰𝗲𝘀 𝗮𝘀 𝗳𝗶𝗿𝘀𝘁‐𝗰𝗹𝗮𝘀𝘀 𝗮𝗿𝘁𝗶𝗳𝗮𝗰𝘁𝘀. If we get this right, we’ll build organizations where agents understand human reasoning, humans understand agent reasoning, decisions become auditable, knowledge compounds, and the 𝗲𝗻𝘁𝗶𝗿𝗲 𝗰𝗼𝗺𝗽𝗮𝗻𝘆 𝗯𝗲𝗰𝗼𝗺𝗲𝘀 𝗺𝗼𝗿𝗲 𝗶𝗻𝘁𝗲𝗹𝗹𝗶𝗴𝗲𝗻𝘁 𝗼𝘃𝗲𝗿 𝘁𝗶𝗺𝗲. That’s the real trillion‑dollar opportunity.... aperturedata.io/resources/cont…

This collaboration also gave us an opportunity to launch UI support for our text vector search features!

I was recently asked about my predictions for vector databases over the next year, and the more I think about it, the clearer one thing feels. 𝗩𝗲𝗰𝘁𝗼𝗿 𝗱𝗮𝘁𝗮𝗯𝗮𝘀𝗲𝘀 𝗮𝗹𝗼𝗻𝗲 𝗵𝗮𝘃𝗲 𝗻𝗲𝘃𝗲𝗿 𝗯𝗲𝗲𝗻 𝗲𝗻𝗼𝘂𝗴𝗵. There is still room to innovate on algorithms and scale, but the bigger shift is how we think about 𝗱𝗮𝘁𝗮 𝗺𝗮𝗻𝗮𝗴𝗲𝗺𝗲𝗻𝘁 𝗮𝘀 𝗮 𝘄𝗵𝗼𝗹𝗲. Moving beyond text-only workloads toward truly 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 data. Treating 𝗺𝗲𝗺𝗼𝗿𝘆 as a first-class concept for the future of AI agents. And questioning whether throwing more compute or long context at the problem is actually the right long-term answer. What’s surprised me most is how little emphasis there’s been on 𝗲𝗳𝗳𝗶𝗰𝗶𝗲𝗻𝗰𝘆. We’ve been comfortable scaling power and cost, instead of demanding better system design and more thoughtful infrastructure choices. Vector search is just one part of the equation. 𝗧𝗵𝗲 𝗿𝗲𝗮𝗹 𝗶𝗻𝗳𝗹𝗲𝗰𝘁𝗶𝗼𝗻 𝗽𝗼𝗶𝗻𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝘄𝗵𝗲𝗻 𝘁𝗲𝗮𝗺𝘀 𝗻𝗲𝗲𝗱 𝘁𝗼 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝘀𝗰𝗮𝗹𝗲, 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗱𝗮𝘁𝗮, 𝗮𝗻𝗱 𝗿𝗲𝗹𝗮𝘁𝗶𝗼𝗻𝘀𝗵𝗶𝗽-𝗮𝘄𝗮𝗿𝗲 𝗾𝘂𝗲𝗿𝗶𝗲𝘀 𝗶𝗻 𝗮 𝘄𝗮𝘆 𝘁𝗵𝗮𝘁 𝘁𝗿𝗮𝗱𝗶𝘁𝗶𝗼𝗻𝗮𝗹 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝘄𝗲𝗿𝗲𝗻’𝘁 𝗱𝗲𝘀𝗶𝗴𝗻𝗲𝗱 𝗳𝗼𝗿. Feels like we are entering a phase where the real differentiation won’t come from models alone, but from how well we design memory, infrastructure, and efficiency into the stack (@ApertureData ). Curious how you are thinking about this as you plan for the year ahead..

𝗔 𝗳𝘂𝗹𝗹 𝗱𝗮𝘆 𝗮𝘁 𝘁𝗵𝗲 𝗣𝗼𝘀𝘁‐𝗜𝗻𝗱𝘂𝘀𝘁𝗿𝗶𝗮𝗹 𝗦𝘂𝗺𝗺𝗶𝘁, 𝗮𝗻𝗱 𝘁𝗵𝗲 𝘀𝗶𝗴𝗻𝗮𝗹 𝘄𝗮𝘀 𝘂𝗻𝗺𝗶𝘀𝘁𝗮𝗸𝗮𝗯𝗹𝗲, 𝗶𝘁'𝘀 𝗻𝗼𝘁 𝗶𝗳 𝘄𝗲 𝗱𝗲𝗽𝗹𝗼𝘆 𝗮𝗴𝗲𝗻𝘁𝘀 𝗶𝗻 𝗲𝘃𝗲𝗿𝘆 𝗼𝗿𝗴 𝗯𝘂𝘁 𝘄𝗵𝗲𝗻 𝗮𝗻𝗱 𝗵𝗼𝘄 𝘀𝘆𝘀𝘁𝗲𝗺𝗮𝘁𝗶𝗰𝗮𝗹𝗹𝘆. (Plus I had the chance to chair the session on Organizational Intelligence, Context Graphs, and Hyper‑Adaptive Enterprises with such insightful speakers like Prukalpa Sankar, Dipanjan Ghosh, Tatyana Mamut) Across conversations, the shift was clear: 𝘞𝘦’𝘳𝘦 𝘮𝘰𝘷𝘪𝘯𝘨 𝘧𝘳𝘰𝘮 “𝘊𝘢𝘯 𝘸𝘦 𝘢𝘶𝘵𝘰𝘮𝘢𝘵𝘦 𝘵𝘩𝘪𝘴?” 𝘵𝘰 “𝘞𝘩𝘢𝘵 𝘥𝘰𝘦𝘴 𝘢𝘶𝘵𝘰𝘮𝘢𝘵𝘪𝘰𝘯 𝘤𝘩𝘢𝘯𝘨𝘦 𝘢𝘣𝘰𝘶𝘵 𝘩𝘰𝘸 𝘸𝘦 𝘸𝘰𝘳𝘬, 𝘨𝘰𝘷𝘦𝘳𝘯, 𝘢𝘯𝘥 𝘤𝘰𝘰𝘳𝘥𝘪𝘯𝘢𝘵𝘦?” 𝗧𝗵𝗲𝗺𝗲𝘀 𝘁𝗵𝗮𝘁 𝗸𝗲𝗽𝘁 𝘀𝘂𝗿𝗳𝗮𝗰𝗶𝗻𝗴 • Everyone is automating something — but the real leverage is in supervisor agents • Context is queen, but culture has to be encoded too • Evals and observability are finally first‑class • As hallucinations drop, the bottleneck shifts to system design and governance • Bounded autonomy + control loops are becoming productivity tools • Cross‑org agent execution? Not yet • Agent sprawl is real • High‑value agents are coming fast 𝗧𝗵𝗲 𝘀𝗽𝗶𝗰𝘆 𝘁𝗮𝗸𝗲𝘀 𝘁𝗵𝗮𝘁 𝗺𝗮𝗱𝗲 𝗽𝗲𝗼𝗽𝗹𝗲 𝘀𝗶𝘁 𝘂𝗽 • “Our teams aren’t allowed to code anymore, they can only train models.” Fascinating, but hard to imagine for infrastructure where correctness and reliability matter as much as intelligence. • “Models evolve, compute gets better, humans struggle to change.” A brutally accurate observation. • “No need to centralize data, let each orgs' agents handle their own. ” Beautiful in theory with the right super agent in charge. In large enterprises with fragmented data estates, this requires deep structural change. 𝘐 𝘵𝘩𝘰𝘶𝘨𝘩𝘵 𝘢𝘣𝘰𝘶𝘵 𝘵𝘩𝘦𝘴𝘦 𝘴𝘱𝘪𝘤𝘺 𝘵𝘢𝘬𝘦𝘴 𝘢𝘯𝘥 𝘵𝘩𝘦 𝘤𝘰𝘮𝘮𝘰𝘯 𝘳𝘶𝘯𝘯𝘪𝘯𝘨 𝘵𝘩𝘦𝘮𝘦 𝘢𝘯𝘥 𝘢𝘭𝘭 𝘵𝘩𝘦𝘴𝘦 𝘱𝘰𝘪𝘯𝘵𝘴 𝘵𝘰𝘰𝘬 𝘮𝘦 𝘣𝘢𝘤𝘬 𝘵𝘰 𝘵𝘩𝘦 𝘦𝘯𝘪𝘨𝘮𝘢 𝘰𝘧 𝘮𝘦𝘮𝘰𝘳𝘺 𝘧𝘰𝘳 𝘢𝘨𝘦𝘯𝘵𝘴! If agents are going to supervise other agents, operate across modalities, and make decisions with bounded autonomy, persistent multimodal memory becomes the real substrate, not just retrieval, but continuity. Follow along on what we are building at @ApertureData to see how we achieve this! Appreciated this opportunity from the Post Industrial Institute team..

** Multimodal Infra FAQs ** Choosing your infrastructure right can save you months if not years of heartache and wasted money. But doing it for AI, especially as you work on building production AI systems, is not easy. Here are some fundamental multimodal infrastructure Q/A that are worth bookmarking as your checklist, if you’re doing AI development beyond tabular data (which incidentally is a modality too!). Q: What is multimodal data? I wish people would ask me that before assuming that if you don’t work with images or videos, you don’t have multimodal data. Nope! Technically, as soon as you start using more than one type or “mode” of data, say text + tables, you are already multimodal. Q: Is there a reason to tie metadata, embeddings, and raw assets together for any reason? Can you query them together? A: Search and retrieval always need you to consider multiple factors. Sometimes you know the exact metadata value to look for e.g. give me all pictures of Scarlett Johannson. But sometimes you only have a vague recollection - show me all movies with “blue aliens”. You want to find your data assets nonetheless. If you don’t stitch them together, the workflow will always be stitched together and hard to scale. And yes, it can be done, with the right infrastructure in place. Q: Why can I not just convert everything to text and use it? A: When you listen to a dialog or watch a scene, is it just words that impact your understanding of the situation or also the tone, the sentiment, and the emotions? These are hard to capture in a transcript. There is also an inherent structure within other modalities e.g. various scenes in a video, diagrams in a document, and so on. Capturing and understanding the modality can give you a much better and correct view of what you are dealing with. Q: Can you actually visualize multimodal data? A: This is quite a valuable capability to be honest. Viewing PDFs, images, and video frames directly speeds up debugging and iteration significantly. And yes, with the right representation and UI, it is possible to do so. Q: What happens when your dataset grows into millions of objects? A: Builders often ask about concurrency, hardware memory behavior, and performance especially under sustained load. Underlying object stores are very well optimized for high throughput. The right representation and parallel compute setup can create a scalable platform to handle multimodal data for AI usage patterns. Q: Can a platform built for multimodal AI support agentic memory requirements? A: Short-term vs long-term retrieval, cross-modality context, and persistence patterns are quickly becoming real requirements and yes, it is possible to build a layer on the right data foundation. These are the kinds of questions that reveal whether a system can support true multimodal AI in production. If you are evaluating infrastructure in this space, these questions are a solid starting point. (check out @ApertureData on how all this is done)

As a data product we had to move fast in 2025, after all, the market required it, moving from vanilla RAG to graph to Agentic RAG in a few months interval each. Now, 2026 is off to a great start at @ApertureData with petabytes of multimodal data, increasing belief in the power of knowledge graphs, and enterprise scale Agentic memory use cases to deploy on the foundation we have built in ApertureDB. Checkout how our year went and what's coming in 2026 : aperturedata.io/resources/refl…

What's also interesting are the lessons around the role of SQL as outlined towards the end of the blog! JSON made functionality more expressible across data modalities while SQL and MCP plug-ins helped offer compatibility...

What went right with AI in 2025? What didn’t get enough attention in 2025 but should — and will — pick up momentum in 2026? This edition of our final newsletter of the year brings together insights from AI leaders across the industry on how adoption matured, why Agentic AI took off, and what must change in 2026 to move from experimentation to impact. Sharing the full state of the union here. linkedin.com/pulse/19-apert… Thank you for your thoughtful responses - Matthias Spycher, Stephanie Cannon, Manasvi Sharma, and Yuanbo Wang!

@ApertureData @ayesha_imr Read the blog here --> aperturedata.io/resources/engi… Try out the agent yourself --> adb-query-agent.netlify.app

In Part 1, @ApertureData valued community member, @ayesha_imr, focused on the data foundation -- designing a multimodal graph in ApertureDB that lets an AI agent understand and traverse conference content using natural language. That structure set the stage for something deeper: 𝗴𝗶𝘃𝗶𝗻𝗴 𝘁𝗵𝗲 𝗮𝗴𝗲𝗻𝘁 𝘁𝗵𝗲 𝘁𝗼𝗼𝗹𝘀 𝗶𝘁 𝗻𝗲𝗲𝗱𝘀 𝘁𝗼 𝗿𝗲𝗮𝘀𝗼𝗻. Part 2 is where she demonstrates how those query patterns evolve into seven well-defined tools, and a LangGraph-based ReAct agent learns how to combine them to interpret real user intent, 𝘁𝗼 𝗵𝗲𝗹𝗽 𝘆𝗼𝘂 𝗻𝗮𝘃𝗶𝗴𝗮𝘁𝗲 𝗰𝗼𝗺𝗽𝗹𝗲𝘅 𝗱𝗮𝘁𝗮 𝗰𝗼𝗹𝗹𝗲𝗰𝘁𝗶𝗼𝗻𝘀 𝗮𝗻𝗱 𝗴𝗮𝘁𝗵𝗲𝗿 𝗶𝗻𝘀𝗶𝗴𝗵𝘁𝘀 𝘆𝗼𝘂 𝗻𝗲𝗲𝗱. In this case, from conference data shared by @MLOpsWorld Here are the top three takeaways from this phase: • 𝘛𝘰𝘰𝘭 𝘥𝘦𝘴𝘪𝘨𝘯 𝘥𝘦𝘵𝘦𝘳𝘮𝘪𝘯𝘦𝘴 𝘢𝘨𝘦𝘯𝘵 𝘤𝘢𝘱𝘢𝘣𝘪𝘭𝘪𝘵𝘺. The seven tools define what the agent can truly accomplish, more than swapping in a larger LLM ever could. • 𝘜𝘯𝘪𝘧𝘪𝘦𝘥 𝘨𝘳𝘢𝘱𝘩 + 𝘷𝘦𝘤𝘵𝘰𝘳 + 𝘮𝘦𝘵𝘢𝘥𝘢𝘵𝘢 𝘴𝘵𝘰𝘳𝘢𝘨𝘦 𝘮𝘢𝘵𝘵𝘦𝘳𝘴. ApertureDB (from ApertureData) enables constrained semantic search (filter--> embed search) in a single atomic query, simplifying the entire workflow. • 𝘎𝘰𝘰𝘥 𝘦𝘹𝘢𝘮𝘱𝘭𝘦𝘴 𝘮𝘢𝘬𝘦 𝘣𝘦𝘵𝘵𝘦𝘳 𝘢𝘨𝘦𝘯𝘵𝘴. Detailed few-shot examples significantly improved tool selection, parameter accuracy, and multi-step reasoning. The result is an AI agent that can answer complex, multi-part questions about years of MLOps conference content with accuracy and context. Blog and agent link in comments below @langchain

As someone who’s often invited to speak at events, and attends my fair share of conferences, I know firsthand how hard it is to find the right talk, the right clip, or even the right year something happened. Slides live in one place, videos in another, transcripts somewhere else. 𝘎𝘳𝘦𝘢𝘵 𝘤𝘰𝘯𝘵𝘦𝘯𝘵 𝘨𝘦𝘵𝘴 𝘤𝘳𝘦𝘢𝘵𝘦𝘥, 𝘣𝘶𝘵 𝘯𝘰𝘵 𝘢𝘭𝘸𝘢𝘺𝘴 𝘴𝘶𝘳𝘧𝘢𝘤𝘦𝘥! That’s the problem we set out to solve with our 𝗻𝗲𝘄 𝗔𝗜 𝗤𝘂𝗲𝗿𝘆 𝗔𝗴𝗲𝗻𝘁. Give it a try lnkd.in/gVNk7cWi and then head over to the blog to read how setup the data layer for it. In this first blog, @ayesha_imr talks about how we partnered with 6th Annual @MLOpsWorld | GenAI Summit committee to build a structured, multimodal data foundation that spans three years of conference content; titles, descriptions, transcripts, speaker info, and videos - all stored inside ApertureDB from @ApertureData. With that architecture in place, the agent can answer questions like: “Which talks covered AI agents with memory?” “Show presentations on RAG by Databricks engineers.” No manual digging. Just natural language --> precise retrieval (isn't that what we all like to do nowadays) This first post in the series focuses on the data layer, the schema and graph design choices that make intelligent query decomposition possible as the effectiveness of any agent is ultimately limited by the structure of its data. Stay tuned for Part 2 where we will explore how these patterns turn into tools inside a LangGraph-based ReAct agent.

ICYMI UT09x02: Moving Beyond Text for Agentic AI Applications with @Vishakha041, founder & CEO of @ApertureData, @FredericVHaren of @HighFens, & @SFoskett of @TechFieldDay #AgenticAI #MultimodelData #AgenticApplications Watch here: buff.ly/5B5rozX

ICYMI UT09x02: Moving Beyond Text for Agentic AI Applications with @Vishakha041, founder & CEO of @ApertureData, @FredericVHaren of @HighFens, & @SFoskett of @TechFieldDay #AgenticAI #MultimodelData #AgenticApplications Watch here: buff.ly/5B5rozX

When you hear “multimodal data for AI,” do you immediately think “embeddings” or hey it doesn’t include my text data? Most people do. But multimodality is so much more than generating embeddings from different data types and running approximate searches. • It’s about 𝘂𝗻𝗱𝗲𝗿𝘀𝘁𝗮𝗻𝗱𝗶𝗻𝗴 𝘀𝘁𝗿𝘂𝗰𝘁𝘂𝗿𝗲: the layout in a PDF, the relationships between fields, the hierarchy of metadata. • It’s about 𝘃𝗶𝘀𝘂𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻: actually seeing the document, the image, the table, the text block you’re working with. • It’s about 𝗻𝗮𝘃𝗶𝗴𝗮𝘁𝗶𝗼𝗻: moving through components, discovering how each element connects to its metadata and to its embedding space. Most importantly, 𝗿𝗲𝗮𝗹-𝘄𝗼𝗿𝗹𝗱 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗱𝗮𝘁𝗮 𝗿𝗮𝗿𝗲𝗹𝘆 𝗳𝗶𝘁𝘀 𝗶𝗻 𝗮 𝘀𝗶𝗻𝗴𝗹𝗲 𝗻𝗼𝗱𝗲’𝘀 𝗺𝗲𝗺𝗼𝗿𝘆. That’s where the real engineering challenges show up. 𝗪𝗲 𝗵𝗮𝘃𝗲 𝗯𝗲𝗲𝗻 𝘁𝗵𝗶𝗻𝗸𝗶𝗻𝗴 𝗱𝗲𝗲𝗽𝗹𝘆 𝗮𝗯𝗼𝘂𝘁 𝘁𝗵𝗶𝘀 𝗮𝘁 @ApertureData 𝗮𝘀 𝘄𝗲 𝘀𝘂𝗽𝗽𝗼𝗿𝘁 𝗣𝗗𝗙 𝘃𝗶𝗲𝘄𝗶𝗻𝗴, 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗽𝗿𝗼𝗰𝗲𝘀𝘀𝗶𝗻𝗴, 𝗮𝗻𝗱 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝘀𝗲𝗮𝗿𝗰𝗵 𝗶𝗻 𝗼𝗻𝗲 𝗽𝗹𝗮𝗰𝗲 𝗮𝗻𝗱 𝘁𝗵𝗶𝘀 𝗱𝗶𝘀𝘁𝗶𝗻𝗰𝘁𝗶𝗼𝗻 𝗸𝗲𝗲𝗽𝘀 𝗰𝗼𝗺𝗶𝗻𝗴 𝘂𝗽. If “multimodal” still feels like “just embeddings,” or only when you have more than text , you are only seeing 20% of the picture.

tldr; A few years ago, I kept seeing the same scene play out in every AI team I met. A room full of brilliant engineers, state-of-the-art models, and GPUs humming in the background, yet half the day was spent moving data between systems. 𝗜𝗺𝗮𝗴𝗲𝘀 𝗶𝗻 𝗼𝗻𝗲 𝘀𝘁𝗼𝗿𝗲, 𝗺𝗲𝘁𝗮𝗱𝗮𝘁𝗮 𝗶𝗻 𝗮𝗻𝗼𝘁𝗵𝗲𝗿, 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀 𝘀𝗼𝗺𝗲𝘄𝗵𝗲𝗿𝗲 𝗲𝗹𝘀𝗲. 𝗘𝗻𝗱𝗹𝗲𝘀𝘀 𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗼𝗿𝘀. 𝗙𝗿𝗮𝗴𝗶𝗹𝗲 𝗽𝗶𝗽𝗲𝗹𝗶𝗻𝗲𝘀. It wasn’t a lack of innovation, it was the architecture. The 𝗺𝗼𝗿𝗲 𝗺𝘂𝗹𝘁𝗶𝗺𝗼𝗱𝗮𝗹 𝗼𝘂𝗿 𝗔𝗜 𝘀𝘆𝘀𝘁𝗲𝗺𝘀 𝗯𝗲𝗰𝗮𝗺𝗲, 𝘁𝗵𝗲 𝗺𝗼𝗿𝗲 𝗱𝗶𝘀𝗰𝗼𝗻𝗻𝗲𝗰𝘁𝗲𝗱 𝘁𝗵𝗲 𝘂𝗻𝗱𝗲𝗿𝗹𝘆𝗶𝗻𝗴 𝗱𝗮𝘁𝗮 𝗴𝗼𝘁. That’s when we started asking a simple question: what if the database itself understood multimodal data? 𝘞𝘩𝘢𝘵 𝘪𝘧 𝘪𝘵 𝘵𝘳𝘦𝘢𝘵𝘦𝘥 documents, 𝘪𝘮𝘢𝘨𝘦𝘴, 𝘷𝘪𝘥𝘦𝘰𝘴, 𝘦𝘮𝘣𝘦𝘥𝘥𝘪𝘯𝘨𝘴, 𝘢𝘯𝘥 𝘮𝘦𝘵𝘢𝘥𝘢𝘵𝘢 𝘢𝘴 𝘧𝘪𝘳𝘴𝘵-𝘤𝘭𝘢𝘴𝘴 𝘤𝘪𝘵𝘪𝘻𝘦𝘯𝘴, not as files to be passed around? That question became 𝗔𝗽𝗲𝗿𝘁𝘂𝗿𝗲𝗗𝗕 (from @ApertureData), a database built for the next era of AI, where agents need to see, connect, and reason in the multimodal world. Intelligence isn’t just retrieval; it is understanding the relationships that tie information together - context through relationships and hierarchies. That’s the mindset guiding every product decision we make. And because I believe in sharing what we learn: • 𝗞𝗲𝗲𝗽 𝗮𝘀𝗸𝗶𝗻𝗴: What relationships do I expect my models/agents to reason over? • 𝗖𝗵𝗮𝗹𝗹𝗲𝗻𝗴𝗲 𝘆𝗼𝘂𝗿𝘀𝗲𝗹𝗳: Does my data layer allow me to query those relationships as easily as I query "select * from …"?

Testing limits of the product and imagining what is possible are what drive a startup in the right direction. At this month’s HackerSquad by Developer Events Hack Day in SF, we saw both happening in real time. Some builders went straight into the technical depth. They wanted to understand how 𝗔𝗽𝗲𝗿𝘁𝘂𝗿𝗲𝗗𝗕 𝗵𝗮𝗻𝗱𝗹𝗲𝘀 𝗾𝘂𝗲𝗿𝘆𝗶𝗻𝗴 𝗺𝗲𝘁𝗮𝗱𝗮𝘁𝗮, 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀, 𝗮𝗻𝗱 𝗿𝗮𝘄 𝗳𝗿𝗮𝗺𝗲𝘀 𝘁𝗼𝗴𝗲𝘁𝗵𝗲𝗿, 𝘄𝗵𝗮𝘁 𝗵𝗮𝗽𝗽𝗲𝗻𝘀 𝗮𝘀 𝗱𝗮𝘁𝗮𝘀𝗲𝘁𝘀 𝘀𝗰𝗮𝗹𝗲, 𝗮𝗻𝗱 𝗵𝗼𝘄 𝗰𝗼𝗻𝗰𝘂𝗿𝗿𝗲𝗻𝗰𝘆, 𝗺𝗲𝗺𝗼𝗿𝘆, 𝗮𝗻𝗱 𝗽𝗲𝗿𝗳𝗼𝗿𝗺𝗮𝗻𝗰𝗲 𝗯𝗲𝗵𝗮𝘃𝗲 𝘂𝗻𝗱𝗲𝗿 𝗿𝗲𝗮𝗹 𝘄𝗼𝗿𝗸𝗹𝗼𝗮𝗱𝘀. These are the kinds of questions that push us to refine and harden the system at @ApertureData Others looked at the same platform and immediately started imagining. Multimodal search across PDFs, images, video, and text sparked ideas around 𝗸𝗻𝗼𝘄𝗹𝗲𝗱𝗴𝗲 𝗴𝗿𝗮𝗽𝗵𝘀, 𝘃𝗶𝗱𝗲𝗼 𝘀𝗲𝗺𝗮𝗻𝘁𝗶𝗰 𝘀𝗲𝗮𝗿𝗰𝗵, 𝗴𝗲𝗻𝗼𝗺𝗶𝗰 𝗲𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝘀, 𝗱𝗮𝘁𝗮𝘀𝗲𝘁 𝗽𝗿𝗲𝗽, 𝗰𝗹𝗼𝘂𝗱 𝗶𝗻𝗴𝗲𝘀𝘁𝗶𝗼𝗻, and more. Seeing all modalities connected in a single system opens doors quickly. 𝘉𝘰𝘵𝘩 𝘢𝘳𝘦 𝘦𝘲𝘶𝘢𝘭𝘭𝘺 𝘪𝘮𝘱𝘰𝘳𝘵𝘢𝘯𝘵. 𝘖𝘯𝘦 𝘬𝘦𝘦𝘱𝘴 𝘶𝘴 𝘩𝘰𝘯𝘦𝘴𝘵. 𝘛𝘩𝘦 𝘰𝘵𝘩𝘦𝘳 𝘬𝘦𝘦𝘱𝘴 𝘶𝘴 𝘪𝘯𝘴𝘱𝘪𝘳𝘦𝘥. Thank you to @itsajchan and the HackerSquad community for bringing together a room full of builders who do both so naturally.