The New Stack’s Post

Here's my new article on vector databases that got published on Towards AI !! #genai #rag #llm #towardsai “The Rise of Vector Databases: Understanding Vector Search and RAG Pipeline“ https://lnkd.in/gpwBJADH What do you think about it?

KNN Search Algorithm Comparison: Optimizing Vector Databases for LLMs and RAG by Danilo Poccia 🔍 Dive into the world of KNN search algorithms! Explore the GitHub project comparing KD-Tree, Ball Tree, HNSW, and Brute Force methods. Uncover insights for optimizing vector databases in AI applications. #KNN #AI #GitHub #VectorDatabases 🤖 https://buff.ly/4djpRf5 #aws #cloudcomputing #devto

🌟 Chatting with SQL🌟 I recently explored the fascinating capabilities of chatting with SQL databases using LLMs, inspired by Tuana Çelik's work with the OpenAI API. 💡 You can check out her insights here: 📊💬 https://lnkd.in/dBgfPMuQ I played with the fine-tuned Mistral version, HuggingFaceH4/zephyr-7b-beta. While it has some limitations compared to the OpenAI API, the potential for natural language querying in databases is truly remarkable. 🚀 #AI #MachineLearning #SQL #NaturalLanguageProcessing #HuggingFace #OpenAI

#LLMs #AI #MonarchMixer #SubQuadratic #GEMM #MachineLearning #MonarchMatrices #BERT #ViT #GPT #GPU #StructuredMatrices ** This blog post from Hazy Research at Stanford University discusses the initial release of their long-context retrieval models with Monarch Mixer, also known as M2-BERT retrieval models. These models are designed to handle long-context retrieval tasks, which are tasks that require understanding and retrieving information from large amounts of text. The models released include M2-BERT-80M-2k, M2-BERT-80M-8k, and M2-BERT-80M-32k, with the numbers indicating the sequence lengths the models can handle. ** The team faced challenges in training these models, as simply extending the standard BERT training pipeline to longer sequence data was insufficient. They found that it was helpful to warm-start the 32K model using the 8K model and copy over positional embeddings at initialization. ** The blog post also shares some performance results. The M2-BERT retrieval models can outperform much larger models on their benchmark, including models up to 4 times larger when fine-tuned and up to 85 times larger for zero-shot models. This suggests that long-context models are beneficial for retrieval tasks' ** The team is seeking feedback from the community on the performance of these models in real-world long-context retrieval applications. They have made the code and models available for download and use via the Together embeddings API. ## Reference: https://lnkd.in/gv6R4QSK https://lnkd.in/geySRTTG

🚀 Exciting Update: Introducing Hybrid Synthesis in IndoxGen! 🚀 We’re thrilled to announce the release of IndoxGen v0.0.9, packed with a game-changing feature: Hybrid Synthesis! 🎉 This powerful new addition combines the best of two worlds: 🔹 Large Language Models (LLMs) to generate realistic and context-aware text data. 🔹 Generative Adversarial Networks (GANs) to produce high-quality tabular data. By merging these two cutting-edge technologies, Hybrid Synthesis allows you to generate both textual and numerical data in a single, cohesive process—perfect for projects that require a diverse and comprehensive synthetic dataset. Key Features of Hybrid Synthesis: ✅ LLM-powered text generation for natural, varied text. ✅ GAN-based tabular data generation, supporting categorical and numerical variables. ✅ Seamless integration into a unified pipeline for balanced and consistent data output. To start using Hybrid Synthesis, simply install IndoxGen along with either indoxGen_torch or indoxGen_tensor based on your setup: pip install indoxgen indoxgen-torch # For PyTorch users or pip install indoxgen indoxgen-tensor # For TensorFlow users Useful Links: GitHub: https://lnkd.in/de7PpuuE PyPI: https://lnkd.in/ddTMtvSy PyPI Torch: https://lnkd.in/diaxsMEy PyPI Tensor: https://lnkd.in/dKPicQrG Whether you’re working on machine learning, data augmentation, or privacy-preserving data sharing, IndoxGen's Hybrid Synthesis opens up endless possibilities for innovation and efficiency. 💡 Upgrade now to IndoxGen v0.0.9 and experience the future of synthetic data generation. 🌐 👉 Learn more and start using IndoxGen today: https://lnkd.in/de7PpuuE #SyntheticData #DataScience #MachineLearning #AI #Innovation #GAN #LLM #IndoxGen #DataGeneration #TechInnovation

Vector Store Start-ups -- do they have a future? Read any book about distributed systems, like Kleppman's classic: "Designing Data Intensive Applications" and you will know the depths, vagaries and nuances of making data-at-scale work. Work on any serious search project and you will also know about the gazzillion nuances, optimization strategies and trade-offs in building a scalable performance search system. And, the moment you go "off piste", which is just about every production project outside of "vanilla search", and the technicalities of building a performant system for a specific set of queries, often with counter-productive constraints, is a non-trivial exercise. One only has to try search on so many products today to see how difficult it is -- try Slack search, try LinkedIn search -- and you will soon get the idea. Go deep into the bowels of Lucene and you will find a treasure-chest of well-honed indexing techniques. Enter vector stores start-ups, many of them opportunistically jumping on the ANN bandwagon and offering ANN-search as a "product". These might be great if all you really want to do is send open-ended Qs to a doc store, but even that doesn't work very well out of the box -- e.g. see the myriad solid case studies from Snorkel AI about getting something to actually work using data-driven optimizations of LLM-backed services: https://lnkd.in/gzqYWVZ8 Moreover, recent work has shown that in many instances, the tried-and-tested BM25 index does better than dense vectors. This might have something to do with the increasing light being shone upon the ineffectiveness of dense vectors to truly encode semantic similarity. One only has to consult the burgeoning field of embeddings-optimization literature to see how the whole "king + man - queen" excitement doesn't hold under many conditions. And, all one should really care about is the actual "X + Y - Z" of the use case, not some idealized benchmark or wow-factor demo. Many critical search queries involve aggregations. These require non-trivial implementations in a distributed system, which is probably why some of the vector stores don't support them. If they do, I suspect they are challenging to tune via any indexing optionality. On the other hand, systems like Lucene and Solr have been in the "real-world" indexing game for a long time and now offer dense vector support. If nothing else, it is worth paying attention to what these communities say and do in terms of technical developments, trying to handle the constraints of multiple-indexing methods in a single platform. Anecdotally, one suspects that they have a better chance of making real-world search work versus start-ups whose only advantage is speed-to-market with an ANN solution. But I remain open-minded of course, as any engineer should. There is no flag-waving in engineering, although I might be old-school in that regard.

Did you miss our Fireside chat on "The Essential Role of Vector Databases in LLMOps" with Sage Elliott and Yujian Tang from Zilliz? Watch the recording or read the transcript! 👇 This AI fireside chat covers: - What is Retrieval Augmented Generation (RAG) for LLMS? - What is LLMOps (Large Language Model Operations) - What is a Vector Database? - Why Are Vector Databases Essential for LLM applications? - What is Multimodal RAG? - What is your LLMOps Developer Stack? - What is Zilliz Building around Vector Databases and LLMs? - What are you looking forward to in the AI Space?     https://lnkd.in/gQeUBYGU

I've just published a new blog post about harnessing the power of GPT for text summarization and categorization! 🤖 💡 Highlights: • Using OpenAI API to analyze and categorize over 500 data science blogs I've read since 2021 • Code example of utilizing Function Calling for consistent API output format • Analysis of my evolving reading interests • Practical industry applications 📖 Check out the blog here: https://lnkd.in/gVheWzYz #chatgpt #gpt #openai #datascience #dataanalytics #dataanalysis #textanalytics #textanalysis

The biggest potential of this tools is as an interdisciplinary tool. Data science is by nature an interdisciplinary area. It joins tools such as machine learning and statistical analysis to other areas, such as medical datasets. When one has a dataset, it is not enough: they need to extract knowledge from that. The approach used will show the knowledge. The coder interpreter also has the ability to suggest what to do with the dataset, using its knowledge as large language model (LLM). For instance, in one case, it was able to guess the meaning of each column from the CSV, and guessed the possible values. The data set was from medicine, from a specialized area. On a real-scenario, that would require a specialized medical doctor, which could be even hard to find around, available for consultation.  #datascience #openai #largelanguagemodels #statisticalanalysis https://lnkd.in/dckSs-CZ

🚀 A new blog on revolutionizing document search using OpenAI's Embedding API and PostgreSQL. Dive into the world of semantic search and discover how AI can enhance data retrieval in your projects! 📊🔍 #OpenAI #PostgreSQL #AI #SemanticSearch #DocumentSearch #TechBlog