Jaron Waldman’s Post

Rerank is the most slept-on model in AI. If you're running RAG pipelines you need to incorporate a reranking step before passing the retrieved docs in the prompt. With Rerank 3, Cohere is enabling a massive step forward in retrieval for RAG.

As always, Cohere gets it right when it comes to answering to enterprise needs.

We just open-sourced one of the more interesting tools we have developed for our design partners - deterministic rule-based retrieval in RAG. Developers frequently tell us that they know exactly where to find the answer to a question within their raw data, but for some reason, their RAG solution is not pulling in the right chunks. To help with this, we are open sourcing a rule-based retrieval solution whereby developers can define rules and map them to a set of chunks they care about, giving them more control in their retrieval workflows. Check it out on Github (https://lnkd.in/e7tzsTQT) and join the conversation on Discord (https://lnkd.in/ezzTkqzj) On our roadmap, we will be linking deterministic rule-based access with knowledge graphs to ensure deterministic semantic retrieval. To get access to this as one of our design partners, ping us at WhyHow.AI. https://lnkd.in/eWaihnsA

Deploying and managing Large Language Models (#LLMs) in production introduces challenges such as ensuring system scalability, optimizing real-time performance, transitioning smoothly from research to production, and integrating a diverse range of data tools and platforms. With Shakudo, you benefit from: ✅ Seamless Transition: From Jupyter Notebooks to production-ready LLM services. ✅ Flexible AI operations: Easily toggle between models like GPT-4, PALM 2, and FALCON LLM. ✅ High-performance vector databases: Ensure consistent and scalable data management, integrated seamlessly with your existing data infrastructure. ✅ No Vendor Lock-in: Over 100 stack components ensure versatility and adaptability. Shakudo addresses all of the key requirements for LLM deployment and management. Read more here: https://lnkd.in/dPDgKssy Ready to optimize your LLM strategy? Schedule a demo: https://lnkd.in/dWkqJtbw

Ready to dive deeper into SearchGPT, our game-changing GenAI-based RAG (Retrieval Augmented Generation) solution? Here are some of the ways it can make a big impact on your business: ⏰ Saves time: Leveraging the power of AI, LLMs, and embedded vector databases, SearchGPT enables lightning-fast, efficient searches throughout your entire document database (and various file types) while upholding the security standards required by your business. 🛠️ Harness the Power of GenAI: With SearchGPT, simplicity meets power. Just input your query and let the GenAI model do the rest, delivering relevant information and document citations instantly. 💻 User-Friendly and Accessible: SearchGPT can be used by employees of all technical levels, ensuring quick and seamless access to the information you need. It's time to revolutionize your document search process. Discover the future of document management with SearchGPT: https://lnkd.in/ezDFVyfF

Ever wondered how embeddings play a crucial role in indexing?👋 These nifty tools help us convert complex data into numerical sets that encapsulate the essence of the information. Think of them as a roadmap in a multidimensional space, where similar words are closer to each other. Now, imagine this for chunks of text in a large document. Consider a massive e-book as your dataset. You want to ask questions about its content, but the document is too vast for an LLM's context window. The solution? Chunk it up!🔍 However, we must ensure these chunks retain the context of the story or characters – no one wants to start reading a book from page 145, right? This is where indexing comes in. Ways to Think About Indexing: 1. Chunk Optimization: 💡Short or long data? Choose your chunking strategy accordingly. 💡Consider your use case: Q&A chatbot, summarization tool, or part of an agentic workflow? 2. Multi-representation Indexing: 💡Parent Document: Retrieve the most related chunk and pass the parent document for improved context. 💡Dense X Retrieval: Propose a new method where propositions outshine sentences and passages in retrieval performance. 3. Specialized Embeddings: 💡Fine-tuning: Improve RAG pipeline performance by finetuning embeddings, offering a 5-10% boost. 💡ColBERT: A BERT-based search model for fast and accurate retrieval over large datasets. 4. Hierarchical Indexing: 💡RAPTOR: Stanford's tree-based model for document summarization, enhancing retrieval accuracy by 20%. Scalable solutions for vector stores like Pinecone, Weaviate, and open-source Chroma are game-changers for large datasets. Whether it's chunking optimization, multi-representation indexing, specialized embeddings, or hierarchical approaches like RAPTOR, each method brings its unique flavor to the table.

Speed Up Your #LLM Applications Using #Semantic Cache Layer. A semantic cache not only stores previous results like a traditional cache, but it also understands the semantic meaning of the query. This means that queries or questions that are not exact matches can still provide previous answers if the intent of the request is the same. In Semantic Caching, we make use of a vector database as a low-latency cache to store all the user question embeddings and the corresponding answers generated in the past to each question. Now, why use SingleStore as the semantic cache layer? SingleStore is a real-time, distributed database designed for blazing fast queries with an architecture that supports a hybrid model for transactional and analytical workloads. This pairs nicely with generative AI use cases as it allows for reading or writing data for both training and real-time tasks — without adding complexity and data movement from multiple products for the same task. SingleStore also has a built-in plancache to speed up subsequent queries with the same plan. Try SingleStore database for free: https://lnkd.in/gfR2nqcU In the rapidly evolving landscape of AI and LLMs, it’s extremely important to ensure that systems are fast, efficient and scalable. A semantic cache layer helps LLM applications to perform with speed. Let's do a simple hands-on to see how it works: https://lnkd.in/g9jKD9Nz ----------------------------------------------------------------- BTW, learn many GenAI concepts in my e-book 'GenAI for Everyone'. Download the 'GenAI for Everyone' E-Book: https://lnkd.in/gy7F3tvB Thanks to Marie Stephen Leo for this amazing workflow image.

Unlocking the power of OLLAMA for LLM: Simplifying complex environment setups for llms with intuitive AI. From streamlining research to enhancing analysis, OLLAMA makes its easier for the user to use large scale llms while it manages the environment #ollama #llm

"The most comprehensive write-up on Indexing", if I may say so myself! Imagine just picking up a book and starting to read from page 145 – how lost for context will you be? Same goes for an LLM trying to understand the context when you feed it only chunks from data that are relevant to the user query. This is what we explore in part 3 of the Advanced RAG series, ie Indexing! This step is crucial in ensuring accuracy, cost efficiency and speed, as we move from the user query transformation to data preprocessing. We will look at various methods such as Chunk Optimization, Multi-representation Indexing, Specialized embeddings and Hierarchical Indexing and go a little deeper into what is the latest and greatest within them. Happy reading! As always, if you like what you see, please feel free to subscribe. I look forward to your comments and feedback! Series inspired by Harrison Chase LangChain Lance Martin #AI #RAG #llmops #indexing #retrievalaugmentedgeneration