Prashant Dixit

Local RAG from Scratch using LLama3 in 5 baby steps Try it - github.com/lancedb/vector… 1. Extract content 2. Recursive Chunking 4. Embed Chunks with @lancedb Embedding API 5. Semantic search with Query, #LLama3 for resulting output using @ollama. Simple Illustration

@GithubProjects My System's RAM 💻

What's stopping YOU from CODING like THIS?

WTF 😱

@ariG23498 @RisingSayak @LysandreJik Yep 👍😊

@Prashant_Dixit0 @RisingSayak @LysandreJik We will be sharing details in some time. The best way to contribute to this (at the moment) would be the join the talks and ask great questions and engage with fellows.

[Hugging Face Machine Learning Club India] We have locked in our first guest for the series. Some updates: 1. The talk is scheduled on 10th of March 2. It is going to be online, (links will be made available) 3. 2000 HRS IST is the set time. We will have more details tomorrow (or the day after). Brownie points for guessing the guest. Let's make it interesting, the title of the talk is "Teaching Tiny Models to Prove Hard Theorems". 🤗

@petergostev What's the current hype?

Do you remember how the hottest thing in AI were vector databases?

@GithubProjects @grok Find the project directory Then do sudo rm -rf project_directory/* PS: This is a versatile command which fixes any bug/issue in your codebase 🥹

Hey @grok what is the most efficient way to get rid of BUGS ? Hint: sudo rm -rf.

@dejavucoder Make designation which doesn't look so middle class 😂

monisha beta, forward deployed engineer bolo. software consultant engineer is too middle class.

Has anyone tried @liquidai 's LFM2.5-1.2B-Thinking Its On-Device Reasoning Model that too under 1GB size, best model to run on mobile phone or low compute devices as it fits in 900 MBs, delivers both the fastest inference speed and the best quality for its size. 🤗Huggingface : huggingface.co/LiquidAI/LFM2.… #thinking #liquidai #mobile

@HarveenChadha @erpriyankpatel I totally agree with you because if we have PHDs but not Infra(GPUs), then what's the sense of having it

@erpriyankpatel Yes, but in 500:1 ratio

can't believe its 2026 and we are still debating if we need GPUs or PhDs

@nikhilaravi @AIatMeta I am using SAM3 for various tasks and even building Automated/ Semi-Automated Labelling platform 😅. Let me know if anyone is up for the collaboration or funding it 😀

@YuXiang_IRVL I would suggest you to refer to the Jetson Model Zoo for some of the already implemented ideas. docs.nvidia.com/tao/tao-toolki…

For a computer vision class project (live demo on Jetson), I currently have: 1. Object detection / tracking 2. Semantic segmentation 3. Hand / human pose estimation 4. Human–object interaction 5. Monocular depth estimation What other demo-friendly ideas would you suggest?

@abhi1thakur Is it still UP or Dead ?, I am interested in becoming a moderator 😅

@abhi1thakur Vibe Coding 😂

wtf is this doing?

New Object Detection dataset is up on @huggingface 🔥SKU-110K - huggingface.co/datasets/Prash… 🫵Follow Dataset collection for upcoming datasets on Huggingface huggingface.co/collections/Pr… @ariG23498 @lhoestq @mervenoyann

@jerryjliu0 Only difference between which I understand activity log and Context graph is that it builds relationship with the information as per the inputs given 📂

I have no idea what a context graph means outside the high-level marketecture Is it just RAG over your activity logs?

@cyberpengk @jerryjliu0 Yes definitely a very large or short both chunk size affects the accuracy, but now it's more adaptive to find the perfect chunking size instead of hard coding it

@jerryjliu0 But "better" size still leads to more accurate embedding? It flips a switch of whether agent can find it at all

yes you still need to chunk "something" if you want to embed content for vector search. my theory is some dumb initial chunking like sentence-level, fixed size, page-level is perfectly fine the agent will jump to a file and navigate around it

@iScienceLuvr New release would definitely pumpup people about @thinkymachines and definitely we can't oversee Soumith, he a gem who made Pytorch. Explore tinker by Thinking Machine thinkingmachines.ai/tinker/

I'm just going to say it: people are being too bearish on Thinking Machines. You guys act like they don't have Soumith Chintala and John Schulman and other really talented folks who I respect a lot. You guys act like they didn't release Tinker, which has inspired similar products at many other companies. You guys act like they don't consistently release high-quality research explorations that serve as valuable contributions to the field.

@craigzLiszt Other competitors are getting created 💁, but Fall of OpenAI doesn't seem quite near. Btw OpenAI is acquiring Thinking Machine Labs founding members back.

are we witnessing the fall of openai?

@jerryjliu0 But still to design a chunking pipeline to be used by an agent to semantically chunking. ✍️I wrote about different chunking techniques before some time lancedb.com/blog/chunking-…

RAG/retrieval might not be dead. But chunking is dead. There is no point overthinking your chunk size if the agent can dynamically expand context around a file.