Jonathan Ellis

Thanks to the resurgence of RL, LLMs are finally able to reliably coordinate tools and reasoning to do high-precision retrieval. Companies like Happenstance, Clado, and Mintlify have already shifted to agentic search, and it's only a matter of time until anything less feels broken to users. Link to full blog in comments.

20

2 Comments

We shipped DeepEval + LangGraph integration. Two lines of code to run eval metrics on your LangGraph agents. People kept telling us LLM-as-a-judge setups require too much boilerplate. Prompt templates, parsing logic, retry handling, score normalization. By the time you're done wiring it together, you've built half a framework. It should be simple. It should be reliable. That's the point of DeepEval. Now with LangGraph: trace your agent, apply metrics, see what breaks. No ceremony. Link below 👇

33

7 Comments

𝗡𝗲𝘁𝗳𝗹𝗶𝘅'𝘀 𝗚𝗿𝗮𝗽𝗵𝗤𝗟 𝗠𝘂𝘁𝗮𝘁𝗶𝗼𝗻 𝗔𝗿𝗰𝗵𝗶𝘁𝗲𝗰𝘁𝘂𝗿𝗲: 𝗗𝘆𝗻𝗮𝗺𝗶𝗰 𝗔𝗣𝗜 𝗜𝗻𝘁𝗲𝗿𝗼𝗽𝗲𝗿𝗮𝗯𝗶𝗹𝗶𝘁𝘆 Their homepage used to be static. Load once, stay unchanged. But users wanted dynamic experiences. Watch comedy trailers, see more comedy content immediately. The challenge was doing this for 200 million users without destroying performance. GraphQL subscriptions would require massive infrastructure investment. Traditional queries cache aggressively, making real-time updates nearly impossible. Polling kills performance. Netflix chose mutations for updates. Sounds wrong since mutations modify server state, but they have one crucial advantage - they never use cached data. Always hit the server for fresh information. The breakthrough was cache normalization. Instead of fetching entire page structures, Netflix separates content into normalized chunks. Page has 100 sections, only 1 is new? Fetch complete payload for the new section, just reference IDs for the existing 99. Data transfer drops 99%. Real innovation is server-driven triggers. Netflix moved update logic from clients to servers. Initial page load includes trigger definitions with encoded action IDs. When events happen, clients pass action IDs back in mutations. Server decides response based on context. Same user action might update a section on mobile, navigate to new page on TV, trigger prefetch on web. All controlled server-side. Netflix can modify user experience across all platforms through configuration. No app updates required. Schema design is elegant. UpdatePageResult contains updated page structure with reference IDs plus separate field with full payloads for new content. Clients merge surgically into normalized caches. Result: sub-second page updates across their entire user base. Netflix can experiment and deploy experience changes instantly across every platform. Netflix proved GraphQL mutations can be the foundation for real-time, cross-platform experiences when combined with intelligent caching and server-driven business logic. You can have both performance and flexibility if you think differently about where intelligence lives in your system. Amazing article by Sreekanth Ramakrishnan, David Shin, Mehmet Mustafa Yilmaz: https://lnkd.in/eD-BmuhR

7

Kafka was founded on its ability to source data from A before sinking it to B,C,D and so on (see Stan’s excellent post here: https://lnkd.in/eTk98Kdr). However, with the rise of Iceberg as the de facto standard for data storage (and associated integration) does this foundational advantage still hold? What is Iceberg? Iceberg is an open table format defining how data is stored and managed in order to provide features that cover the most common needs of data focused applications including:  ⚫  Schema evolution  ⚫  Partitioning  ⚫  Time Travel More importantly, it won the “table format wars” and has become the fastest growing storage format for data systems. Iceberg is engine agnostic, meaning that any number of processors, data lakes and database providers are jumping on the Iceberg bandwagon and providing support for Iceberg as base storage. What is a Kafka Sink? A Kafka Sink is a component that takes data from Kafka, accumulates/formats it for a downstream system before writing to it. The most popular tool for this is Kafka Connect, a distributed framework that allows high performance, fault tolerant data pipelines to be created and managed with no code. The pre-Iceberg pattern is for a single copy of data to be written to multiple downstream systems using Sinks. You may use the S3 Sink to write to your lake whilst at the same time use a JDBC Sink to write to the database that powers your dashboards. It’s not immediately apparent how Iceberg changes this, isn’t Iceberg not just another destination for our sink connector? Yes and no, the difference is that Iceberg is itself a flexible and well supported data source that can power our applications. There are a number of advantages to this approach:  ⚫  𝗥𝗲𝗱𝘂𝗰𝗲𝗱 𝘀𝘁𝗼𝗿𝗮𝗴𝗲 𝗿𝗲𝗾𝘂𝗶𝗿𝗲𝗺𝗲𝗻𝘁 - Instead of a copy for each downstream system only one copy needs to be maintained.  ⚫ 𝗔 𝗰𝗼𝗻𝘀𝗶𝘀𝘁𝗲𝗻𝘁 𝗳𝗼𝗿𝗺𝗮𝘁 and set of capabilities - Sinks downstream system dependent. Iceberg provides a consistent view that can be relied upon by all clients.  ⚫  𝗡𝗼 𝗿𝗮𝗰𝗲 𝗰𝗼𝗻𝗱𝗶𝘁𝗶𝗼𝗻 between sinks - Each sink is independent and this can lead to races (one sink may have processed data that our another has not). Iceberg ensures consistency by being the sole source.   ⚫  𝗙𝗮𝘀𝘁𝗲𝗿 𝗮𝗱𝗼𝗽𝘁𝗶𝗼𝗻 - Any Iceberg compatible downstream system can instantly use the existing Iceberg data available. A Sink based approach must first populate an initial data set. In light of Iceberg’s growing dominance and its ability to act as a single, authoritative source for a wide array of systems, the traditional Kafka Sink pattern begins to show its limitations. There’s no silver bullet however, (remember that streaming Iceberg has its own problems: https://lnkd.in/eq3vgnzJ) but it’s a powerful new weapon in the armoury of modern data infra.

182

17 Comments

MCP/A2A: I’m seeing a new wave of Apache Iceberg articles that mostly re-explain the basics—evidence that adoption is spreading. Alex Merced has launched an LLM/MCP series, and Dremio just unveiled Smart Iceberg Table Management. Roy Hasson and Russell Spitzer, however, haven’t posted much on Iceberg lately. With Iceberg likely past the Gartner Hype Cycle spotlight, the bigger opportunity now is to double down on MCP and A2A. comments?

10

1 Comment

Parsing dense tables from documents is challenging - 1. Models miss cells, or mess up the structure of tables. 2. VLMs have limits in number of output tokens, so dense tables are often truncated. 3. Merged headers with multiple rows result in misplaced cells. Financial Services and Research Assistant platforms building AI products rely on tables in documents, so this was an important problem for us to solve. Antaripa Saha wrote a comprehensive blog on how to handle dense tables using Tensorlake! You can go from a Table to Markdown, HTML or data-frames, summarize the table so AI agents can retrieve tables based on what they contain, not just raw text. https://lnkd.in/gnKDS8tt

10

OpenAI has launched Open Responses — a new open API standard (based on their own Responses API) designed to unify how we interact with open models for agentic workloads. The backstory: While OpenAI's Responses API hasn't seen widespread adoption for other API providers, the Claude/Anthropic API has become the hot standard — largely thanks to Claude Code. Chinese model providers like Kimi K2 and Moonshot AI have even stood up Claude API endpoints for their models! What Open Responses brings to the table: 🛠️ Standardized tool calling across open models 📡 Built-in streaming & multimodal input support 🤖 Native agentic loop support for running agents server-side 🔌 Multi-provider by default with extensibility for custom features 📦 Major adopters: Hugging Face, Ollama, Vercel, Open Router, LM Studio, vLLM 💻 Code works today: Hugging Face inference endpoints ✅ Ollama local models ✅ Reasoning trace streaming with GPT-OSS models ✅ 🔮 2026 Prediction: Expect Frontier open model labs (Qwen, Kimi, DeepSeek) to train models compatible with BOTH Open Responses AND the Anthropic API. Ollama just announced Anthropic API compatibility too — meaning high-quality local models running with the ability to use Claude Code tools is not too far away. This could be a huge win for developers wanting to switch between proprietary and open models without rewriting their stack. 🚀 Check out the video and demo code in the

82

2 Comments

Now is the time to migrate from #ApacheSolr to #OpenSearch For years, Apache Solr has been a reliable search workhorse - but the search landscape is evolving fast. With Solr’s development cadence slowing and community momentum thinning, teams aiming to modernize their search stack are increasingly looking to OpenSearch Project instead. Here’s why forward-thinking engineering leaders are making the switch: - Modern Search Capabilities Built-In: OpenSearch delivers vector search, hybrid semantic + lexical search, and native machine learning integrations - enabling richer, AI-powered search experiences that users expect today. - Cloud-Native & Fully Managed Options: With Amazon OpenSearch Service and other managed offerings, your team can stop worrying about cluster ops, upgrades, and maintenance and focus on delivering more value. - Rich Analytics & Observability: OpenSearch Dashboards offers integrated visibility into search performance and system health - without stitching together third-party tools. - Better Total Cost of Ownership: Even if self-managed Solr looks cheaper up front, the operational overhead - infrastructure, tuning, bug fixes - adds up. OpenSearch often wins the ROI game when you factor in people and time. Yes - migration isn’t a one-to-one lift-and-shift: you’ll map schemas to OpenSearch mappings, translate query logic to its Query DSL, and adapt configurations. But it's also not that big of a deal. We just completed a few migrations like that. With careful planning, parallel environments, and a strong proof-of-concept, you can mitigate risk and unlock future growth. Migration guide is in the comments :) Let’s modernize search — and get more value for every buck invested. 💡

59

7 Comments

Interesting article on the state of vector databases. Tracks well with the evolution of TxtAI. At one point along the way, TxtAI focused on being a vector database. Then it evolved into being more of an orchestration framework like LangChain with a solid foundation of an embeddings database (sparse/keyword, dense, graph components) https://lnkd.in/e7sn96-S

13

4 Comments

The hidden thing that decides how much a GPU costs. Why chips from the same batch can have totally different prices 👇 I recently learned about something called "Chip Binning", and it turns out it’s a key reason why there are so many versions of the same CPU or GPU. And why some cost more than others. It’s not just a tech term, but it actually affects how much we pay and how much performance we get. 🟣 So what Is "Binning"? When companies make computer chips, they produce hundreds from a single round piece of silicon called a wafer. After making them, each chip is tested to see how well it performs (how fast it runs, how hot it gets, and how stable it is). Based on these results, chips are sorted into categories, or “bins.” The best-performing integrated circuits are placed in the top bin and sold as high-end models (like an RTX 4090 or Ryzen 9). Chips that don’t perform quite as well still work fine, but they become mid-range or budget models. That’s how the same basic chip design can end up being sold at very different prices. 🟢 Why Binning Is Important Binning helps manufacturers avoid wasting materials. Even if a chip isn’t perfect, instead of throwing it away, they can disable the weaker parts and sell it as a lower-tier product. For example, if a chip was supposed to have eight working cores but only six pass the quality test, they’ll sell it as a six-core version. This way, nearly every chip that works even partially can still be used. 🔵 When Companies Intentionally “Downgrade” ICs Sometimes, the opposite happens: the production run is so good that most chips meet top-tier standards. When that happens, companies will intentionally disable some working parts to create mid-range models. This is called "artificial binning". It’s not about performance issues, it’s just a way to keep a full range of products at different price points. 🟠 Binning in Graphics Cards (GPUs) Binning works the same way for graphics cards. A GPU that doesn’t quite meet RTX 3080 standards might become an RTX 3070 or 3060 instead. 🟡 The “Silicon Lottery” This is a popular phrase among industry experts. It basically means that even two identical-looking CPUs or GPUs can perform differently, depending on how well they were binned. You might buy the same model as someone else but get one that overclocks better or stays cooler. It’s partly luck. 🟤 How Binning Affects Price and Performance Top-binned chips cost more because they’re guaranteed to perform better, use less power, and run cooler. However, I understood that the performance difference between a mid-tier and top-tier chip isn’t always that big. Sometimes, paying extra just gives you a small boost in speed or a fancier cooling system. Binning is both a smart engineering process and a marketing tool. It helps manufacturers make the most of their materials while giving buyers multiple options at different price levels.

28

2 Comments

Talked to a founder on Monday that swore codex was better than claude. They had 11 MCPs configured in claude that chewed up over 50% of the context window. Their codex setup had 0 MCPs. Their claude config was like telling someone about railroads, Japan, and Apple then asking them a calculus problem.

7

2 Comments