Madan Ram’s Post

Check out our latest post on Clappy.ai, "Leverage Layout-Aware RAG for More Accurate Retrieval of Knowledge from Documents." Discover how to harness the power of Retrieval-Augmented Generation (RAG) for documents in PDF, Doc, and Markdown formats. Dive into the specifics of implementation and explore our reference repository to get started with advanced RAG for your knowledge base. Don't miss out on this essential read! Continue reading here for all the details: The Power of Hierarchical Indexing with RAG. (https://lnkd.in/gbiVHfcH)

𝙋𝙖𝙥𝙚𝙧 : 𝙎𝙚𝙖𝙧𝙘𝙝𝙞𝙣𝙜 𝙛𝙤𝙧 𝘽𝙚𝙨𝙩 𝙋𝙧𝙖𝙘𝙩𝙞𝙘𝙚𝙨 𝙞𝙣 𝙍𝙚𝙩𝙧𝙞𝙚𝙫𝙖𝙡-𝘼𝙪𝙜𝙢𝙚𝙣𝙩𝙚𝙙 𝙂𝙚𝙣𝙚𝙧𝙖𝙩𝙞𝙤𝙣 - 𝙐𝙨𝙚𝙧 𝙌𝙪𝙚𝙧𝙮 𝘾𝙡𝙖𝙨𝙨𝙞𝙛𝙞𝙘𝙖𝙩𝙞𝙤𝙣: Decides whether the user query necessitates retrieval or can be answered based on the LLM's general knowledge. It affects latency. The authors used BERT-base-multilingual to classify between 15 task queries. - 𝘿𝙤𝙘𝙪𝙢𝙚𝙣𝙩 𝘾𝙝𝙪𝙣𝙠𝙞𝙣𝙜/𝘾𝙝𝙪𝙣𝙠 𝙎𝙞𝙯𝙚: Determining the chunk size is very important. They use LlamaIndex evaluation modules to compute faithfulness and relevancy over a set of chunk sizes to select the optimal one. - 𝙀𝙢𝙗𝙚𝙙𝙙𝙞𝙣𝙜 𝙈𝙤𝙙𝙚𝙡: Recommending LLM-Embedder, though it depends on the use case and knowledge base. - 𝙑𝙚𝙘𝙩𝙤𝙧 𝘿𝙖𝙩𝙖𝙗𝙖𝙨𝙚: Recommending Milvus based on key criteria: multiple index types, billion-scale vector support, hybrid search, and cloud-native capabilities. - 𝙍𝙚𝙩𝙧𝙞𝙚𝙫𝙖𝙡 𝙈𝙚𝙩𝙝𝙤𝙙𝙨: Recommending Hybrid Search with HyDE. - 𝘿𝙤𝙘𝙪𝙢𝙚𝙣𝙩 𝙍𝙚𝙧𝙖𝙣𝙠𝙞𝙣𝙜: Recommending DLM Reranking using monoT5 reranker. They also recommend best practices for document repacking and summarization of retrieval results. Most importantly, they highlight the use cases of Multimodal RAG, specifically text2image and image2text retrieval capabilities. Paper: https://lnkd.in/eiFK8_CG

Finally, we have it! After one month of tremendous effort, in the frame of D4R project we developed an implementation of the article "PromptORE - A Novel Approach Towards Fully Unsupervised Relation Extraction." Link at the end of the post. We called our creature "Spanish_PromptORE" We use BERT-like models to obtain relationships among textual already extracted entities. If you are looking for a way to measure the branding and engagement of your business, here you have it! The model allows you to extract relations and classify then based on XML-TIE inputs with already extracted entities. Link to the repository: https://lnkd.in/eSuCm6aA Link to the articule: https://lnkd.in/e_Gf5Tp8

😫 Running into context limits when summarizing text using LLMs? Here's a trick for when you run into that issue: 1) Divide the text into multiple chunks. 2) Summarize each chunk. 3) Combine the summaries. If you still run into context limits, repeat steps 1 and 2. 4) Generate a summary of the combined summaries. This divide-and-conquer approach is called recursive summarization. The number of chunks depends on the desired summary detail, the model's context size, and the text length. If you're going to use this approach, I suggest writing a helper function to determine if a chunk size will exceed context limits, given the model and text length. Then, use a while loop to increment the number of chunks until it doesn't exceed the limits. It'll be a while until infinite context is here (although I have a feeling that statement will age like milk), so for now, recursive summarization is a decent substitute.

Nice read (and image) by Leonie Monigatti about hyperparameter optimization for RAG applications (this is when your LLM can read stuff before it answers). What would you like to optimize in your RAG applications? https://lnkd.in/enCAucwh

Text classification is the process of categorizing or labeling text data into predefined categories based on its content, enabling the organization and analysis of textual information. In the following course, we cover the basics of text classification: https://lnkd.in/dWtcpirs

We're happy to share our best-seller, ‘The Ultimate Guide to Vector Databases’! 📚 This guide has become popular on CTO bookshelves, as many realize that building RAGs is a complex process that can go wrong in many ways: 👉 Document retrieval  👉 Text generation  👉 RAG scaling in production Start your chapter on vector databases today: https://lnkd.in/eQKTyJS2

MASSIVE AMOUNTS OF KNOWLEDGE YET IT’S SOMEWHAT HARD TO CONSUME. There’s a lot of TL;DR including probably this write up! The problem: ——————— Indeed we are in the “Information Age”. We’ve got Wikipedia, Brittanica, Wikimedia, Google's sunset product - Knol, and the various News websites that serve valuable information and news upon querying them. The proposal: ——————— We already have LLM based systems that summarize any complex articles from these sources. But these are manual systems - so far, presumably every article and diagram on Wikipedia was created by humans. How would the world look like if we typed a question and an expert system constructed an easy to grasp explanation in real-time - with words, pictures, diagrams, flowcharts, tables, audio, and video clips, as necessary? A 4th grade level explanation when a 4th grader asks - simple pictures and brief sentences, and a COO level elaboration when that's appropriate - with detailed and current pie charts, tables and reports etc - all without the user realizing or needing to demand it. The solution: ——————- “Generate context appropriate output from any type of input query for any type of context”... The current LLMs trained on trillions of tokens, with many orders of magnitude more data than humans ever see in their lifetimes, still don’t have as vivid of an understanding or explainability of the causal relationships that exist in the human world. I am working on a small writeup to demonstrate my understanding of a possible solution to this. Stay tuned for now ... because one day that writeup itself could be autogenerated and may have slight variations, depending on who's the reader!

[pdf] Foundations Of Decision-Making Agents: Logic, Probability and Modality Subrata Das digsell https://lnkd.in/erbgfWCb This self-contained book provides three fundamental and generic approaches (logical, probabilistic, and modal) to representing and reasoning with agent epistemic states, specifically in the context of decision making. Each of these approaches can be applied to the construction of intelligent software agents for making decisions, thereby creating computational foundations for decision-making agents. In addition, the book introduces a formal integration of the three approaches into a single unified approach that combines the advantages of all the approaches. Finally, the symbolic argumentation approach to decision making developed in this book, combining logic and probability, offers several advantages over the traditional approach … Read More » https://lnkd.in/eHvgieqg