Li Yin’s Post

It seems like nowadays, no product can afford to be without AI/LLM-based capabilities, whether those capabilities bring actual value to their users or are used purely from a marketing perspective. So, we're experiencing an LLM-boom. Everyone seems to be RAG-ing, fine-tuning, or even training models from scratch. This reality requires software architects and engineers, who may not have years of in-depth knowledge in the AI/ML field, to design, develop, and deliver production-ready LLM (or AI, if you will) capabilities. So, I thought I'd start a "for dummies" (from a dummy :-)) series of entries that cover just enough knowledge to deliver these capabilities in a timely manner, but not so in-depth as to include knowledge you will not necessarily use. First 2 kick-off entries: LLM Fine Tuning for Dummies: https://lnkd.in/dDmxnGBN Quantization and Post-Quantization Performance Assessment for Dummies: https://lnkd.in/djj_kHPU

How to analyze data using plain English translated to SQL with AI. Inspired by Harshad S's Na2SQL application, I built a Retrieval Augmented Generation(RAG) application powered by LlamaIndex to analyze US Government contract data using plain English. With this application, I demonstrate how to use custom datasets to ask data questions in natural language and return the corresponding SQL(Structured Query Language) statement in addition to the expected results using LlamaIndex and Openai. To learn more, check out my GitHub application: Video demo: https://lnkd.in/eWxaHa49

GPT-4 On-the-Fly DataVisualization Tools: A Simple Mapping Example https://bit.ly/3VjQ5c7

Retrieval-Augmented Generation (RAG) vs Fine-tuning: A Quick Dive! 🚀 🔍 Introduction: Navigating the world of Large Language Models (LLMs)? Two popular optimization techniques stand out: RAG and Fine-tuning. Each shines in its unique way. 🎯 Fine-tuning: What: Customize LLMs for specific tasks using niche datasets. Example: Making an LLM write e-commerce product descriptions? Fine-tune with existing descriptions! Visualization: Data ➡️ Model Initialization ➡️ Fine-tuning ➡️ Evaluate ➡️ Deploy! 🌌 RAG: What: Enhance LLMs by pulling external data, perfect for tasks like Q&A. Visualization: User Query ➡️ Vector Encoding ➡️ Search Vector DB ➡️ Retrieve Document ➡️ LLM Response ➡️ User gets the answer. ✅ How to Choose: Needs: RAG for tasks demanding external data. Fine-tuning for internalizing knowledge. Constraints: Got massive labeled datasets? Fine-tuning is your friend. Short on it? RAG might be the answer. 💰 Costs: Fine-tuning can be affordable, but watch out for model size and dataset intricacies. Some strategies, like Low-Rank Adaptation (LoRA), can save the day! 🔥 In a Nutshell: RAG = Efficient Data Retrieval Fine-tuning = Deep Learning & Pattern Recognition For top-tier LLM optimization, weigh your app's needs and pick wisely. Sometimes, a hybrid can bring the magic! #MachineLearning #RAG #FineTuning #DeepLearning #AI

GPT-4 On-the-Fly DataVisualization Tools: A Simple Mapping Example https://bit.ly/3VjQ5c7

How To Make LLMs Master Tabular Data? The challenge of processing and understanding tabular data has been a long-standing hurdle. But the tide is turning with the groundbreaking "Chain-of-Table" method! This innovative approach is transforming how LLMs interact with tables, treating them like dynamic puzzles to be solved piece by piece. It mirrors the human method of thinking, allowing the model to iteratively update a table, reflecting each step of the thought process. Imagine an AI's cognitive journey unfolding in a structured, spreadsheet-like format – that's what Chain-of-Table offers. The results are already making waves, with Chain-of-Table achieving unprecedented success in some of the most demanding table reasoning challenges like WikiTQ, FeTaQA, and TabFact. 👉 https://lnkd.in/ecmq_hxM Stay tuned as we continue to explore and expand the possibilities of AI and data interaction! #AIInnovation #DataRevolution #ChainOfTable #LLMs #TechTrends

The last method we explored for the use of LLMs with custom data is RAG. RAG stands for "Retrieval-Augmented Generation," a powerful method that combines information retrieval and language generation to enhance the capabilities of LLMs. We're making LLMs smarter by adding more helpful information to the questions we ask. It's similar to providing extra clues, so LLMs can give really good answers. Let's talk about the practical steps to set up RAG for your specific needs. Step 1: Preparation To get started, we need a database containing the information we want the large language model to work with. This could include web pages, PDFs, books, internal documentation, or any data relevant to your queries. Step 2: Splitting Data It's crucial to split the data thoughtfully, perhaps by paragraph or section, to maintain context. This ensures that when inserted into the model, you get comprehensive information, allowing the model to provide more accurate responses. Step 3: Insertion and Embedding Insert the split data into your database. Alongside the text, convert it into an embedding – a vector format. This mathematical structure enables the model to compare vectors, judging their similarity. We use an AI model for generating these embeddings, storing both the text and embeddings in our database, such as vectordb. Step 4: Retrieving Information When it's time to call the large language model, and a user inquiry comes in, we use the same model to create an embedding from the inquiry. Leveraging vector search, we compare distances between vectors in our database to find the most relevant results. These results are then used in our prompt. Step 5: LLM Generates Output The large language model generates an output incorporating our custom data, providing users with accurate and tailored information. RAG is an exciting tool that, when set up correctly, enables dynamic knowledge retrieval, allowing your large language model to shine in delivering precise responses. #LLM #dataengineering

🚀 Excited to share my latest article on Medium! Learn how to easily enhance RAG performance and token saving using LLM rerank and similarity postprocessors. RAG, while powerful, often retrieves chunks with high scores but lacks relevant context. Our solution? Implementing two post-processing steps to refine the results. 🔍 Semantic search optimization: We fine-tuned parameters and applied a similarity cutoff to filter irrelevant chunks, ensuring a more focused context. 🔄 LLM rerank with context optimization: By reranking and selecting only the most relevant chunks, we significantly improved response quality and reduced token usage. Check out the full article for insights on our methodology and results: 👉 https://lnkd.in/dTwwNnaj Don't miss this effective approach for optimizing RAG performance with LlamaIndex! #AI #RAG #LLM #SemanticSearch #Optimization #Lognext

Imagine that you can ask questions to your data directly. With combination of #generatieveai and #pandas as well #openai you can talk to your #dataframe using #pandasai ! And combining PandasAI with Large Language Models (LLMs), will offer powerful tool for intuitive data interaction and cleaning, allowing users to command data more naturally. OpenAI's API powers PandasAI, simplifying dataset analysis within notebooks. The blend of LLM's contextual understanding with the structured data analysis makes this integration particularly insightful, paving the way for deeper insights and enhanced efficiency. One example of LLM modeling with PandasAI: https://lnkd.in/dxFM3-7B