Isabel González’s Post

In the realm of AI, embeddings play a pivotal role by transforming non-numeric data into numerical form, enabling computers to process text and images effectively. In a Simple Guide to Retrieval Augmented Generation, you'll read about - - The significance of embeddings in data science, machine learning, and AI - Various embedding algorithms such as Word2Vec, GloVe, and BERT - Practical examples of using embeddings to enhance text search, clustering, and more Advancements in Large Language Models (LLMs) have revolutionized industries worldwide, but with great power comes unique challenges. As we marvel at their text generation capabilities, concerns about reliability and trust have surfaced. To tackle these issues head-on, I've delved into the innovative world of Retrieval Augmented Generation (RAG). Having spent years developing the production-grade LLM-based application Yarnit, RAG has been instrumental in my journey. I'm thrilled to announce my upcoming book on RAG with Manning Publications Co., a comprehensive guide tailored for technology professionals eager to harness the potential of LLMs. 📚 What You'll Learn: - Understand the fundamentals of Retrieval Augmented Generation - Build LLM-based applications from scratch - Navigate RAG pipelines from design to production 🔍 Chapters Include: - Large Language Models and the Need for RAG - Designing RAG-Enabled Systems - Creating Knowledge Bases with Indexing Pipelines - Real-Time Interaction and Contextual Responses - Evaluating RAG Performance - Technological Foundations for RAGOps - Evolution of RAG Systems: From Basic to Advanced - Nuances: Comparing RAG with Fine-Tuning and More - Cutting-Edge Best Practices and Further Exploration 🌟 Manning Early Access Program (MEAP) Benefits: - Immediate access to the book's draft and future updates - Opportunity to shape the final content with your feedback - Exclusive 50% discount until July 4th, 2024, with code "mlkimothi" Join me in exploring the forefront of AI innovation and mastering RAG. Let's navigate the complexities together and unlock the full potential of Large Language Models. 👉 Reserve your copy today and embark on your journey with RAG: https://lnkd.in/gHkBVW8W Thank you for your support and feedback as we dive into this transformative field. Feel free to drop your thoughts or questions about Manning Early Access Program (MEAP) below. Let's build the future of AI together! #AI #LLMs #RAG #MachineLearning #ManningPublications #MEAP #TechnologyInnovation #Techbook

In the realm of AI, embeddings play a pivotal role by transforming non-numeric data into numerical form, enabling computers to process text and images effectively.

Generative AI has been a game changer in increasing our efficiency. One such application of Gen AI is Retrieval-Augmented Generation (RAG), which is revolutionizing how we integrate search and text generation using LLMs with our own data. I have created a step by step guide on building RAG with LangChain and Google's Gemini Pro and Pinecone . This guide will teach you how to create a basic RAG with chunking, embedding and vector database storage and retrieval. This can be used for summarizing research documents or smaller text based pdfs. Check out the guide on medium and Google colab notebook below : Medium Blog : https://lnkd.in/drET4H7B Google colab Link : https://lnkd.in/dNXiXihV Github Repo : https://lnkd.in/dYXyvT6E #GenerativeAI #RAG

🚀 Unlocking the Power of Retrieval-Augmented Generation (RAG) 🤖 In the rapidly evolving landscape of AI, RAG is emerging as a game-changer! 🌟 🔍 What is RAG? RAG (Retrieval-Augmented Generation) is a powerful hybrid model that enhances traditional LLMs (Large Language Models) by integrating a retrieval system. It combines the strengths of both search-based and generative AI to deliver highly accurate and contextually relevant responses. 📚🔗 📊 How Does It Work? 1️⃣ Document Processing: Knowledge sources (PDFs, images, CSVs) are chunked and embedded for AI understanding. 🗂️🧠 2️⃣ Semantic Cache & Vector DB: Data is stored and indexed for quick retrieval. 🗄️💾 3️⃣ Backend API & LLM: Queries are enriched with context and processed by an LLM for accurate responses. 🛠️📝 💡 Why RAG? Precision: Delivers context-aware, accurate responses. 🎯 Scalability: Handles massive data with ease. 🌐 Versatility: Ideal for domains like customer support, research, and more. 🛡️ ✨ RAG is not just about generating answers—it's about generating the right answers by retrieving and understanding context. RAG is leading the next wave of AI innovation. Excited to see where it takes us! 🚀 #AI #MachineLearning #DeepLearning #ArtificialIntelligence #DataScience #TechInnovation #RAG #LLM #FutureTech #Innovation

🚀 Day 3 of the 5 Days of Gen AI Intensive by Google – learning to build interactive AI applications with databases and function calling! Here’s what I accomplished today: 💬 Developed a system to talk directly to a database using function calling, enhancing automated information retrieval and interaction. 🤖 Built an agentic ordering system in LangGraph, which can process queries and dynamically respond based on structured data! 💡 What I Learned: How function calling can streamline database interactions with natural language. Designing an AI system that autonomously navigates tasks like ordering and answering complex queries. Leveraging LangGraph for structured, agent-based AI systems. Check out my progress on GitHub and Kaggle: 📂 GitHub: https://lnkd.in/g8wdW6wy 📊 Kaggle build langGraph : https://lnkd.in/ga4y9dA2 📊 Kaggle Talk to a database with function calling : https://lnkd.in/g-G42qbx #GenerativeAI #AI #MachineLearning #5DaysOfGenAI #Google #LangGraph #Kaggle #GitHub #AIApplications #ArtificialIntelligence #LearningJourney

🚀 𝗔𝗜 𝗪𝗮𝗿𝘀: 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸, 𝗢𝗽𝗲𝗻𝗔𝗜, 𝗚𝗲𝗺𝗶𝗻𝗶, 𝗟𝗹𝗮𝗺𝗮 & 𝗦𝗟𝗠𝘀 The AI arena is buzzing with DeepSeek-V3, GPT-4o, Gemini 2.0, Llama 3.3, and the dark horse—SLMs (Small Language Models). Let’s break it down (with humor, because AI needs laughs too):   --- 🔍 𝗗𝗲𝗲𝗽𝗦𝗲𝗲𝗸-𝗩𝟯 The open-source underdog. 685B params (only 37B active/token) + $5.5M training cost = coding/math beast. Think: “𝗔𝗰𝗲𝗱 𝘁𝗵𝗲 𝗲𝘅𝗮𝗺 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗯𝘂𝘆𝗶𝗻𝗴 𝘁𝗵𝗲 𝘁𝗲𝘅𝘁𝗯𝗼𝗼𝗸.” 💸 𝗚𝗣𝗧-𝟰𝗼   The Ivy League diva. Costs $2.50M input/$10M output tokens—like paying for caviar at a diner. Brilliant but pricey.   🎥 𝗚𝗲𝗺𝗶𝗻𝗶 𝟮.𝟬   Multimodal maverick. Handles text, images, video. 1M token memory = AI with a photographic brain. Still experimental—Google’s *“𝘁𝗿𝘂𝘀𝘁 𝗺𝗲, 𝗯𝗿𝗼”* energy.   🦙 𝗟𝗹𝗮𝗺𝗮 𝟯.𝟯   Open-source OG. Dependable, multilingual, but resource-heavy. The tech equivalent of a gas-guzzling SUV.   ⚡ 𝗦𝗟𝗠𝘀   The minimalist rebels. Small, fast, task-specific. Like espresso shots—no fluff, all focus. Perfect for edge computing. *𝗟𝗲𝘀𝘀 𝗶𝘀 𝗺𝗼𝗿𝗲*.   --- 🏆 Verdict   - DeepSeek: Open-source MVP.   - GPT-4o: Premium, pricey, powerful.   - Gemini: Flashy but unproven.   - Llama: Reliable workhorse.   - SLMs: Efficiency kings.  Winner? Open-source + innovation. The future is collaborative, efficient, and (please) cheaper. ---   #AI #DeepSeek #OpenAI #Gemini #Llama #SLM #MachineLearning #OpenSource P.S. Still using GPT-3.5? Your AI deserves glow-up. 😉

Attended a sharing by Gabriel Chua on creating a daily AI paper summariser at last week's ML SG meetup. Here's my take on it: https://lnkd.in/gKWXPC2i Building on top of Gabriel's work, with addition features (keeping with the spirit of the project, all the output was clobbered together from #Claude, #GPT-4o, Co-pilot and #Gemini): - Auto-classification (currently based on HF taxonomy) - Curation (select the topics of interest to display in the readme) - Parsed Huggingface and Github links (not super accurate, sometimes it takes from other references in the paper) - Single page viewer with the results nicely formatted (The first attempt from Claude was very good out of the box, but it took 15+ tries and more specific prompts + some console inspection work to get all the condition styles and logic right) Not too bad for a half day of work

DocGenome: Unlocking a new era in Document AI 🌍📄 I recently came across DocGenome, an incredible open large-scale scientific document dataset designed for training and evaluating multi-modal large language models (LLMs). It’s an impressive resource that could reshape how we approach document understanding tasks. 💡 What Makes DocGenome Stand Out: Completeness: The dataset structures information from all modalities, including 13 layout attributes and their \LaTeX\ source codes. Logicality: Captures 6 logical relationships among document entities for deeper analysis. Diversity: Supports a wide range of tasks—document classification, layout detection, visual grounding, QA, and more. Correctness: Rigorously quality-checked by experts to ensure accuracy. 📊 Key Features: 500K annotated scientific documents from 153 disciplines in the open-access arXiv community. 6.8M pages in the training set and 9K pages in the test set. Multi-level relationships between document components, like "subordinate" and "explicitly-referred." Comprehensive coverage of fixed and floating-form units (text, equations, tables, figures). 🔥 Why It Matters: DocGenome empowers researchers to tackle multi-modal reasoning and layout-aware tasks with high-quality structured annotations. It’s a unique blend of breadth, depth, and attention to detail in document AI datasets. 📂 Explore the Dataset: Available on <https://lnkd.in/g2RnWDFj> Paper on arXiv: https://lnkd.in/gqAp6HgQ GitHub Repository: UniModal4Reasoning/DocGenome I’m genuinely excited about the possibilities this dataset opens up for Document AI research and applications. Whether it’s improving document classification, enabling multi-page QA, or advancing visual-textual reasoning, DocGenome seems to be a game-changer. #DocumentAI #MachineLearning #AIResearch #DocGenome #MultiModalLLMs

🚀 Journey into Generative AI with Fine-Tuned DistilGPT-2 and Wikipedia Data 🚀 I'm excited to share a project I recently worked on, leveraging the power of DistilGPT-2 and Wikipedia data to create a question-and-answer system! 📚🤖 In this project, I: 1. Scraped Wikipedia for specific topics like world history. 2. Loaded the data into my DistilGPT-2 model. 3. Fine-tuned the model with this data, allowing it to answer related questions. This hands-on experience was incredibly valuable and taught me a lot about the intricacies of fine-tuning language models. I used Google Colab for running my code, which was a great tool but did have some GPU limitations. Despite these constraints, I managed to create a functional system and learned a ton along the way! 💡 Key learnings included: - Understanding training arguments such as warmup steps, learning rate, training batch size, and epochs. - Gaining insights into the fine-tuning process and its impact on model performance. This project showcases my commitment to the field of generative AI and LLMs. I look forward to exploring more advanced models and overcoming hardware limitations to push the boundaries of what's possible. #AI #MachineLearning #GenerativeAI #NaturalLanguageProcessing #DistilGPT2 #FineTuning #GoogleColab #AIProjects

The first three chapters for A Simple Guide To Retrieval Augmented Generation published by Manning Publications Co. is now out for early access. This book is aimed at tech professionals looking to get an understanding of RAG in building LLM based applications. From introducing the technique to building RAG pipelines in production, in 9 chapters the book covers the following - Chapter 1 - Large Language Models and the Need for Retrieval Augmented Generation Chapter 2 - RAG-enabled systems and their design Chapter 3 - Indexing Pipeline : Creating a knowledge base for RAG based applications Chapter 4 - Generation Pipeline: Real time interaction for contextual responses Chapter 5 - RAG Evaluation : Checking accuracy, relevance and faithfulness Chapter 6 - Evolving RAGOps Stack : Technologies that make RAG possible Chapter 7 - Progression of RAG systems : Naive to Advanced to Modular Chapter 8 - Nuances : Comparison with fine-tuning, multimodal and agentic RAG Chapter 9 - Cutting Edge : Best practices and further exploration As a precursor, attached are the excerpts from the book covering key ideas conveyed in the first three chapters. As part of the Manning Early Access Program (MEAP) you get access to the live book and the discussion forum where you can give me your suggestions and comments on the chapters as they release. You can subscribe to both the e-book and the print copy which will be delivered to you once the book is complete in the next few months. Grab your copy at a discount today here - https://lnkd.in/gHkBVW8W Use coupon code 'mlkimothi' at the checkout for a 50% discount. I am eager to hear your suggestions to make the book a great resource for anyone interested in building generative AI into their application. #rag #llm #generativeai #artificialintelligence