Eduardo Ordax’s Post

What is an agentic LLM workflow? This is an agentic LLM workflow.

After Retrieval-augmented Generation (RAG) and Fine-tuning (FT) -- which Pascal Biese defines as two techniques for adapting models to the practical realities of real-world data -- are you ready to discover Retrieval-augmented Fine Tuning (RAFT)? The article is worth reading but I am missing principled design, which are embedded in a number of smaller specialist large language models. Techniques to get more out of generalist LLMs yield smaller incremental progress than combining ontologies, knowledge graphs and domain classification with large language models specially designed to perform a limited number of tasks well.

Good piece on the enterprise evolution of LLM implementation from Cobus Greyling. Models morphing, tools exploding, then folks realize the data challenge. Back to my list of helpers: IBM Dell Technologies, Hewlett Packard Enterprise, Databricks MongoDB SingleStore Salesforce, Oracle SAP Pure Storage and more

Excited to share a recent project that highlights the importance of optimization in data processing and its connection to the broader field of data science and AI! I started with a complex and unoptimized script designed to fetch search volume data from an API for various keywords across multiple states. The initial version of the code was far from ideal—it was slow and cost-inefficient due to redundant API requests and poor handling of large data sets. Realizing the need for improvement, I embarked on a journey to optimize the script. The key steps included: 1. **Batch Processing**: Instead of sending API requests for each keyword individually, I grouped keywords into batches, significantly reducing the number of API calls. 2. **State-Specific Processing**: I refined the script to handle only the specific states required, streamlining the data processing workflow. 3. **Error Handling and Logging**: Added robust error handling and logging to ensure that any issues were promptly identified and addressed, allowing the script to continue running smoothly. These optimizations, though simple, had a profound impact. The revamped script not only reduced our data processing costs but also cut down the time required to complete the task drastically. What would have taken two people a considerable amount of manual effort was accomplished in just 1/6th of the time. This project is a great example of how data science techniques, including data collection, processing, and optimization, play a crucial role in AI and machine learning workflows. Efficient data handling is foundational to training robust AI models, making these optimizations essential. Special thanks to my team for their support throughout this process and to our General Manager, Vikram sir, for his invaluable guidance and encouragement. This experience underscored a valuable lesson: sometimes, a few thoughtful improvements can lead to significant gains in productivity and efficiency. Optimization is not just about speed—it's about working smarter, not harder! #DataScience #AI #Optimization #Efficiency #Innovation #TeamSupport

Before 5 months, when I explored to customize a foundational LLM, most of the recommendations were pivoted around fine-tuning a LLM using the domain specific knowledge. Of course, there are a lot of articles explaining the process, but I realized it is expensive to fine-tune a LLM while balancing the cost, performance, and accuracy. It's also not easy to avoid the costly consequences like hallucination (with source clarity), access restrictions with leaving knowledge at inference time, fine-tune repetitions with latest data. I'm glad to see pivotal advances around RAG (Retrieval-Augmented Generation) based LLMs, showing signs to avoid some of the costly consequences from fine-tuning process. Some key benefits include: corroboration without hallucination, maintainable search index referencing latest data, clear source indication to general vs knowledge, access restriction based on license tiers etc. RAG-based solution also accelerate launching your AI applications must faster while leveraging LLMs for your domain specific knowledge base. The Data Management and Vector Databases in this article showcases important tools to help with RAG-based LLMs. https://lnkd.in/gxgjsHNQ

🚀 Excited to share my latest Medium article on fine-tuning LLM applications using LangChain! 📝 In this piece, I delve into some of the key tools in the LangChain framework that users can leverage to enhance the accuracy of their applications' responses. #LangChain #LanguageModels #AI #Technology

𝕴𝖓𝖙𝖗𝖔𝖉𝖚𝖈𝖎𝖓𝖌 Verba 1.0: 𝕬 𝕲𝖆𝖒𝖊-𝕮𝖍𝖆𝖓𝖌𝖊𝖗 𝖎𝖓 𝕴𝖓𝖋𝖔𝖗𝖒𝖆𝖙𝖎𝖔𝖓 𝕽𝖊𝖙𝖗𝖎𝖊𝖛𝖆𝖑 𝖆𝖓𝖉 𝕽𝖊𝖘𝖕𝖔𝖓𝖘𝖊 𝕲𝖊𝖓𝖊𝖗𝖆𝖙𝖎𝖔𝖓 𝑲𝒆𝒚 𝑯𝒊𝒈𝒉𝒍𝒊𝒈𝒉𝒕𝒔: - 𝐇𝐲𝐛𝐫𝐢𝐝 𝐒𝐞𝐚𝐫𝐜𝐡 𝐚𝐧𝐝 𝐒𝐞𝐦𝐚𝐧𝐭𝐢𝐜 𝐂𝐚𝐜𝐡𝐢𝐧𝐠: Combines semantic and keyword search for faster, more accurate data retrieval. - 𝐄𝐧𝐡𝐚𝐧𝐜𝐞𝐝 𝐐𝐮𝐞𝐫𝐲 𝐏𝐫𝐞𝐜𝐢𝐬𝐢𝐨𝐧: Improved ability to handle complex queries and diverse data formats. - 𝗦𝘂𝗽𝗽𝗼𝗿𝘁 𝗳𝗼𝗿 𝗠𝘂𝗹𝘁𝗶𝗽𝗹𝗲 𝗗𝗮𝘁𝗮 𝗙𝗼𝗿𝗺𝗮𝘁𝘀: Efficiently processes various data types like PDFs, CSVs. - 𝐀𝐝𝐯𝐚𝐧𝐜𝐞𝐝 𝐑𝐀𝐆 𝐓𝐞𝐜𝐡𝐧𝐢𝐪𝐮𝐞𝐬: Applies filters and autocompletion before performing Retrieval-Augmented Generation. - 𝐏𝐞𝐫𝐟𝐨𝐫𝐦𝐚𝐧𝐜𝐞 𝐌𝐞𝐭𝐫𝐢𝐜𝐬: Significant improvements in query precision and response accuracy. Discover how Verba 1.0 is transforming AI applications with its innovative approach and robust performance. Sources https://lnkd.in/gpPBv5Xy https://lnkd.in/gxCnHPgE https://lnkd.in/geYTtMd3 Connect with me for more insights!

I want to explain Retrieval as if you are an 8th grader. Not an expert at RAGs? This is for you Types of Retrievals/Searches using AI: >> Text Search (OG Method) Search for pieces of information in your data that have most (if not all) of the same words as your query. This is simple and fast but might miss some relevant information. >> Semantic Search Look for pieces of information in your data that are similar to your query based on their English meaning. This helps find more relevant results even if the exact words don’t match. >> Filter + Semantic Search First, the system figures out some characteristics (metadata) to filter your data. Then, it looks for pieces of information that are similar to your query based on their English meaning. This adds an extra layer of accuracy. >> Summary Search Search for pieces of information based on summaries that have been pre-computed for your data. This method is cheaper than searching over the whole dataset but can be inaccurate for certain tasks. >> Ensemble Search Combine all the techniques above and use either the union (everything that matches any method) or intersection (only what matches all methods) of the results. This ensures you get the best possible results from your searches. Ensemble is the technique that most ML engineers use to win Kaggle ;) This is for people who are not spending hours and hours studying AI engineering. It misses some nuances on these various types, but it gives you a bird's eye view of things. What is the common retrieval mode you see for most AI apps? #aiengineering

Elevate your models with #FineTune! 🚀 Customize pre-trained models for better performance using your data. Spend less time training and achieve superior results. Learn more in our #KnowledgeBase article. 👇 #aiXplain #AI #AIoptimization #AItools

[#𝟭𝟮] 𝗬𝗼𝘂 𝗖𝗮𝗻'𝘁 𝗠𝗮𝗻𝗮𝗴𝗲 𝗪𝗵𝗮𝘁 𝗬𝗼𝘂 𝗖𝗮𝗻'𝘁 𝗠𝗲𝗮𝘀𝘂𝗿𝗲 Measurable doesn't necessarily lead to manageable. I agree with Don Peppers's perspective, which actually isn't contradictory to Peter Drucker's stance. As the question is: 𝗰𝗮𝗻 𝘄𝗲 𝗺𝗮𝗻𝗮𝗴𝗲 𝘄𝗶𝘁𝗵𝗼𝘂𝘁 𝗺𝗲𝗮𝘀𝘂𝗿𝗶𝗻𝗴? Data science is an empirical science. Each model built is the result of an experiment. Given the multitude of potential configurations across data, parameters, and algorithms, we conduct hundreds, even thousands of experiments, seeking the optimal combination for the best performance. But how can we achieve this without measurement, or more importantly, without robust tracking? 𝗧𝗿𝗮𝗰𝗸𝗶𝗻𝗴 𝗶𝘀 𝘁𝗵𝗲 𝗽𝗿𝗼𝗰𝗲𝘀𝘀 𝗼𝗳 𝗿𝗲𝗰𝗼𝗿𝗱𝗶𝗻𝗴 𝗺𝗲𝗮𝘀𝘂𝗿𝗲𝗺𝗲𝗻𝘁𝘀, 𝗮𝗻𝗱 𝗽𝘂𝘁𝘁𝗶𝗻𝗴 𝘁𝗵𝗲𝗺 𝘁𝗼𝗴𝗲𝘁𝗵𝗲𝗿, 𝘀𝗼 𝗮𝘀 𝘁𝗼 𝗴𝘂𝗶𝗱𝗲 𝘂𝘀 𝗳𝗼𝗿 𝘀𝘆𝘀𝘁𝗲𝗺 𝗮𝗱𝗷𝘂𝘀𝘁𝗺𝗲𝗻𝘁𝘀, 𝗮𝗻𝗱 𝗱𝗲𝘁𝗲𝗿𝗺𝗶𝗻𝗶𝗻𝗴 𝘁𝗵𝗲 𝗻𝗲𝘅𝘁 𝘀𝘁𝗲𝗽𝘀. In my ML model development, I actually use two experiment tracking systems. One is a personal setup, which is based on the configuration classes that are saved after each training and testing (as JSON files); the other is an established experiment tracking tool, like #MLflow. For those starting in machine learning modeling, setting up an experiment tracking system early is invaluable.  Here is a comprehensive overview of the top 15 ML Experiment Tracking tools: https://lnkd.in/euiWU_4h, provided by #neptune. If you're yet to decide on yours, explore these options and test them out. #machinelearning #artificialintelligence #datascience #ml #ai