Someshwaran Mohankumar

@saen_dev @mdancho84 Nowadays, for NER, the transformers are doing a much better job at scale. And, in terms of PDF, it became a default training set. Can expect better performance. Will try this LangExtract and will share some results.

@mdancho84 Open source document extraction that actually works would kill a lot of enterprise pricing. The real test is how it handles scanned PDFs with mixed layouts because that's where every extraction tool falls apart in production.

RIP document extractors. Google just released LangExtract: Open-source. Free. Better than $100K enterprise tools. Here’s what it does: 🧵

@TheSuperEng @drummatick When DeepSeekv3 launched, the West strategically moved to restrict NVDA chip exports from the US as it became very powerful. If similar progress is now happening through Chinese alternatives for Mythos, just imagine the impact 😂

@drummatick This is what I said earlier lol. Wait until they open source mythos level AI.

"DeepSeek is a threat." "It's a security concern." "It's Chinese." Dude, go ahead and ask American companies to do more open source. We have cheaper models, free access to open weights, and don't need to rely fully on monopolies, thanks to Open weight models.

I think part of the reason we got into this situation is the tooling. The AI ecosystems like LangChain, LangGraph, LlamaIndex, and AutoGen (and more) made it easy to build agentic RAG systems. So we as devs, pushed to a thought process, what if, if the pipeline we build today need flexibility and scale in the future, even when a simple RAG + function call would’ve been enough. Feels like using a sledgehammer to crack a nut.

Was discussing this with a friend building RAG pipelines. Not every problem needs Agentic RAG, even simple function calls are enough. Feels like the K8s-on-a-resume situation: powerful, but it's an overkill for an "Hello World" application. No need to burn tokens when you know the answer is right in front of you! 😄

This is the most common question I’ve been asked recently: “Should I replace my RAG system with an agentic RAG one?” Well, are you happy with your current RAG system’s performance? Then, probably not. The key difference between traditional RAG and agentic RAG is HOW the context is built: Traditional RAG has a fixed retrieval pipeline that retrieves exactly once. → The LLM is a passive recipient of additional context Agentic RAG has access to one or more retrieval tool(s) to retrieve on demand. → The agent actively builds its own context > Does your use case always require exactly one retrieval step? (Example: always pull in customer information) > Or do you have cases where no additional context is needed? > Or do you have cases where multi-hop retrieval is necessary? If you don’t know, set up a simple evaluation with a gold test set of a few common user patterns: > Do precision and recall improve for your retrieval component? > How much does the end-to-end performance improve based on performance improvements of the retrieval component? > Do these performance improvements justify the added latency and cost of agentic RAG?

@drummatick I've been falling back and forth with this thought process off-late, to use SKILLS.md or just using the function or tool call directly that saves my time. I completely agree with you. Given there's no abstraction needed for all minimal efforts or routine tasks.

My prediction is, slowly everyone will realize that using SKILLS to implement a predictable workflow like browsing, checking logs, or testing use just wasting useless tokens and just having the raw code is better than have SKILLS.md later converted into the code

@vavanessadev @Varnika99 @parthfullstack If you have more control over the memory layer, you can index those episodic memories to a persistent layer like an index, create a connector/tool/MCP to sync between Claude and Codex. This got more success when you switch. Kinda of vendor neutral 😅

@vavanessadev @Varnika99 @parthfullstack I agree with the thought process here, the main reason why we should do this is to persist the memory layer basically filesystems like AGENTS.md, CLAUDE.md with the context to the Codex so it works better.

Why is everyone shifting from claude codex? Should I also consider it?

Exploring a local-first agent setup with continue.dev + Ollama + Gemma4:26b (MoE) with M4 Pro, 48GB. Any suggestions before I go deeper?

@aravindputrevu Yeh toh irony hai! 😂 Technology kitni bhi aage jaaye, containers ho, cloud ho, ya ab LLMs, andar se wahi AWK, SED, aur Bash chal raha hai. Linux fundamentals ek baar seekh lo, zindagi bhar kaam aate hain. Low-level kabhi nahi badalta, sirf uske upar ka layer badalta hai.

Our Unix professor in college made us write AWK scripts as part of coursework. We felt that who needs this today, lol? We thought we would end up writing in a new-age programming language like Python. Fast-forward to 2026 - staring at LLMs writing scripts in SED and AWK 🙂 > echo PRESENT

@zan2434 @code_rgb @eddiejiao_obj @drewocarr Love it. Looking forward to understand the behind the scenes.

@code_rgb @eddiejiao_obj @drewocarr will put out a more in depth explanation on how it works in a few days! but long story short a lot of activation caching, quantization, and torch.compile + memory snapshotting

Imagine every pixel on your screen, streamed live directly from a model. No HTML, no layout engine, no code. Just exactly what you want to see. @eddiejiao_obj, @drewocarr and I built a prototype to see how this could actually work, and set out to make it real. We're calling it Flipbook. (1/5)

@zan2434 @eddiejiao_obj @drewocarr Does a new query always reset the state? Trying to figure out if we can persist state within the same session or if it’s all just temporary caching. If I clear a query and run it again, does the browser cache pick it up? To avoid gen. twice the same query!

@zan2434 @eddiejiao_obj @drewocarr This is a wonderful representation how Text > Image > Video > Interactive pixels. I see when I enter a query, it creates draft version as an Image and then the model goes into turning it into a video.

@the_mcnaveen I have a Q, what's the platform framework you built this on? Flutter or React native?

> builds android app > hit $100 > now build ios app > millionaire soon

@the_mcnaveen Why NOT? 🍀

@dadoonet @xeraa I guess, David was carrying the ELKy same place, last time? 😃

@xeraa I love this place 😍

took some time off in preparation for las vegas; and #GoogleCloud — busy week ahead but looking forward to seeing lots of people

Most agent knowledge bases go stale faster than people realize. If you’re periodically re-embedding docs, your retrieval layer is already drifting out of date. Elastic Workflows fix this with a live context pipeline: Trigger → crawl (Tavily, Firecrawl) → transform → index → retrieve at runtime Your agents always query fresh context instead of snapshots.

@elastic AutoOps (for @elastic_devs), it'll change the way you diagnose your self-managed clusters (tbh, it does). Now, you can connect your clusters for in-depth visibility and anomalies! DIY - spin up a cluster using /start-local.sh and connect AutoOps to see for yourself! 🔥

@shouryagupta132 @elastic @HashiCorp @elastic_devs Thanks for the shoutout! Appreciate it.

Last Saturday was so much fun! Was at the Meetup @elastic x @HashiCorp User Group BLR. Deep-dive talks, electric energy ⚡️ @som23x @elastic_devs thanks for such a great event!

Beyond #ELK in Chennai ✅ #Elasticsearch + ML + RAG + ES|QL + Alerting, all in practice. Strong discussions around grounding, LLM integrations & o11y and ML. #Elastic #RAG #DevOps #OpenTelemetry

We were genuinely surprised 🫣 ...by how interactive it was! Meetup @elastic x @HashiCorp UG BLR. Talks on RAG travel chatbot, Terraform CI/CD, OTEL + HashiCorp O11y (@zachjonesnoel) , and Quantization. Packed. Tough QnA. Serious energy. 🎥 here: youtu.be/UrZOccm8Qjw