Ramanathan Parameshwaran

Perhaps the most critical thing in a Retrieval Augmented Generation (RAG) system is the retrieval or searching strategy. Poor retrieval will lead to irrelevant documents going to the LLM leading to irrelevant generated answers or worse hallucinations. Is cosine similarity everything? Do check out below illustration which talks about four of the most common strategies for playing around with retrieval strategies in RAG systems: 1. Similarity-based retrieval: This is based on embedding similarity, you can play around with different metrics like cosine, maximal margin relevance etc. 2. Similarity with Threshold retrieval: Useful where we cap the similarity threshold and can help in removing irrelevant document but can drop recall 3. Multi-query retrieval: Make the LLM generate variations of your query and get documents based on these multi-versions of your initial query, often useful to get more recall or coverage of documents 4. Reranker retrieval: Use a model specialised in ranking documents based on search relevance Often you can combine some of these techniques and even use ensemble retrievers. Do explore them and go beyond just cosine similarity!

17

Routing the control flow within an RAG application based on the user’s query intent can help us create more useful and powerful Retrieval Augmented Generation (RAG) based applications. #llms #rag #chatgpt4 #gemini #aiml #llmops https://lnkd.in/d54eVrHZ

8

Join me for an insightful session on May 10th, 2024, where I'll be sharing how Games24x7 has leveraged Databricks to simplify the creation and deployment of ML models. I'll discuss our approach to automating MLOps pipelines and developing a solution supporting practitioners and managers throughout the ML lifecycle. From data preprocessing to model training, deployment, and monitoring, we'll explore how automation and AI can streamline these processes, freeing up time for experimentation and innovation. Looking forward to having you join us for this engaging discussion! #ML #AI #Databricks #Games24x7 Registration link: https://lnkd.in/gNTvMw_D

70

2 Comments

Excited to share that I had the honor of being a guest speaker on the latest episode of "The Fluid Intelligence Podcast" by Factspan ! 🎙️ In this episode, we dive deep into the technical challenges and triumphs of building enterprise GenAI applications. It's packed with insights and practical advice for anyone navigating the GenAI landscape. A big thank you to Factspan for this opportunity! 🙏 Stay tuned for the release and don't miss out on this informative session. #Podcast #GenAI #DataScience #DataAnalytics #AI #factspan

17

1 Comment

📍 💡 Need some visual inspiration? #Recommendations to foster new ideas and explore the world. #Recsperts Episode No. 22 with Prabhat Agarwal and Aayush Mudgal from Pinterest is out now! Prabhat is a Senior ML Engineer and Aayush is a Staff ML Engineer. Prabhat works on #recommendations and #search systems at Pinterest, leading representation learning efforts. Aayush is responsible for #ads #ranking and privacy-aware conversion modeling. In our interview, Prabhat takes us on the journey of #personalizing the Pinterest Homefeed. We discuss how to capture users' short- and long-term preferences through multiple #embeddings and the role of candidate generators for content diversification. Aayush dives into the specifics of ads ranking at Pinterest and helps us to better understand how multifaceted ads can be. He also shares some insights about #exploration and corresponding randomization in the context of ads and how user behavior is very different between different kinds of ads. Both highlight the role of #counterfactual evaluation and its impact for faster #experimentation. ✅ Subscribe on your favorite podcast player ⭐ Leave a rating or review ✍ Share your thoughts on Pinterest Homefeed Personalization and Ads Ranking Listen to this new episode of #Recsperts - Recommender Systems Experts! 🎧 Website: https://lnkd.in/gVP5Ar_B 🎧 Spotify: https://lnkd.in/gPyguxY3 🎧 Apple: https://lnkd.in/gDCJjFzw

47

5 Comments

As someone deeply involved in Natural Language Processing (NLP), I cannot stress enough the importance of embeddings and data in any Large Language Model (LLM). Let's not underestimate the significance of these components. Understanding them can provide a competitive edge in the NLP industry special with systems like RAG. In addition, tokenizers play a vital role in languages with special cases like Arabic. Proper handling of tokenizing Arabic text is essential. One of the best practices I've discovered in my career is to separate Prefixes, Affixes, and Suffixes. This approach allows us to unlock a wealth of knowledge and insights that undoubtedly enhance our careers in the legal field. Seeing a lack of attention given to these crucial components is disheartening. I'd like to ask all those interested in LLMs to take the time to understand and appreciate their significance. Don't miss out on the opportunity to gain a deeper understanding of NLP and the competitive edge it can provide. #NLP #LanguageModel #Tokenization #ArabicLanguage #Tokenizers #LLM

33

3 Comments

Bridging Causal Inference and AI Interpretability: Insights from Patchscopes In exploring the Patchscopes framework, I was intrigued by its resemblance to the do-operation in causal inference. Patchscopes performs interventions on a model's internal representations, akin to setting a controlled condition to observe outcomes, much like a do-operation tests causal relationships by manipulating variables. This approach not only enhances understanding of deep neural networks but also bridges concepts between model interpretability and causal inference, offering a robust tool for uncovering how models process information. This insight could significantly enrich our strategies for analyzing complex AI systems. https://lnkd.in/eM6g2mhg

4

Want to perform hyperparameter tuning on your ML and DL algorithms? Learn about how you can do it with Optuna and neptune.ai

15

Building effective RAG Systems by Stanford : https://lnkd.in/gBrTGyci Retrieval Augmented Generation (RAG) systems are being widely adopted to enhance the accuracy and customisation of LLM responses on private data sources. However, most of the industry currently uses "Frozen RAG system", which combines pre-trained and off-the shelf retrieval and language model components without further optimisation. While simple and cost-effective, this approach has limitations such as relying solely on in-context learning, leading to potential hallucinations and factual inaccuracies. The stanford lecture shares ongoing research areas on developing better RAG systems where the retriever and generator are trained and optimized together. The future of RAG lies in creating end-to-end systems that balance cost and quality, rather than simply combining pre-trained components. #RAG #RetrievalAugmentedGeneration #AI

2

Want to know how RL systems can achieve optimal policies. Discover the essential strategies behind defining successful RL systems in my latest blog. Dive in: https://lnkd.in/g4RbkBBi

6

I was reading about Phi-2 as I have been interested in making large language models smaller. This is the blog that I read: https://lnkd.in/gwvikUhZ Question now is very simple, can this be the rise of so called small language models? Should we not try to make large language models smarter but with without having that mountain of redundancy of large language models?

1

A clear explanation of the Stable Cascade Architecture.

4

Do we really need complicated and complex infrastructure to build an impactful AI application? 🤔 Well, check out our blog on how to address one of the practical pain points in #data dissemination—the #discoverability of data—using simple and low-resource #AI. Our goal in this post is to show that there are “low-hanging fruit problems” that can be solved or at least mitigated by AI by leveraging open-source solutions such as the #GISTEmbed model, Hugging Face #transformers.js, and #SentenceTransformers. We demonstrate this by showing how #semantic search over embedding vectors can overcome some of the limitations of traditional #keyword-based systems. But… this is not just a blog post—you can try and compare semantic search and keyword search over The World Bank - World Development Indicators (#WDI) on the blog itself! And… did I say that the search app on the blog is running fully on your browser? No vector #embedding server, no #GPU server for inference, no #cloud storage, etc. We are bringing the model to your browser! This is what small and performant embedding models can unlock—a useful and delightful AI experience on the #edge! :) Imagine having these kinds of edge-driven, low-resource AI applications implemented by countries and organizations with no computing infrastructure and limited resources. I am excited about how far we can enable developing countries and organizations with solutions like this to make their data more accessible and discoverable! We also discussed why query parsing is important to improve the precision of results and maximize the value of metadata. I am also excited about dynamic and context-aware query expansion. I hope that LLM models will become smaller and faster while maintaining performance to make this solution more practical for prime time. Blog: https://lnkd.in/e9yttBvt GitHub: https://lnkd.in/eViwqKAB P.S. Larger embedding models still outperform small models, so there’s a trade-off between resource needs and performance. cc: Olivier Dupriez #AI #developmentdata #semanticsearch #keywordsearch #metadata #lowresourceAI #edgeAI #AIforData #DataforAI #embeddingmodels #opensource

38

2 Comments

Krish Naik recommended this fantastic blog about Transformers, and I stumbled upon it while learning about NLP. The blog breaks down how Transformers function, from attention techniques to their role in speeding up tasks. If you're interested in NLP, this blog is essential reading. Make sure you don't miss it! https://lnkd.in/gDKhwHFy #Knowledgesharing #learningjourney #datascientist #NLP #Transformers #AI #opentowork

27

2 Comments

🚀 Excited to share my latest blog on evaluating the performance of Large Language Models (LLMs)! 🚀 In the realm of NLP, LLMs are revolutionizing how enterprises leverage data, but deploying them in real-world settings poses significant challenges. My new article delves into various evaluation techniques and metrics crucial for assessing LLM performance, with a focus on Retrieval-Augmented Generation (RAG) use cases. Key Highlights: 🔹 Automatic evaluation metrics: Exact Match, BLEU, ROUGE, and Word Error Rate (WER). 🔹 Human evaluation and the importance of platforms like Argilla for feedback. 🔹 RAG systems evaluation using metrics like Recall@k, MRR, and NDCG@k. 🔹 Model-based evaluation and benchmarks: TruthfulQA, MMLU, Big Bench, GLUE, HELM. 🔹 Hybrid approaches combining human judgment with model-based evaluation. Check out the full article for a deeper understanding of these techniques and their impact on improving LLMs in real-world applications. Looking forward to your thoughts and feedback! Let's continue advancing the dynamic field of NLP together. 💡 #AI #NLP #MachineLearning #LargeLanguageModels #RAG #EvaluationMetrics #HumanEvaluation #DataScience #TechInnovation #RedHat #RedHatEmergingTechnologies

13

2 Comments

The quantized version of a language model to perform text generation tasks in both Hindi and English. It's fine tuned aditi model on latest dataset release by sarvamai. https://lnkd.in/gvQ55Vd8 and other hindi datasets. Try out it out free on colab https://lnkd.in/gvbmBzSA model link: https://lnkd.in/gDKqSwtq. #huggingface #datasets #tuningmodel #ML #machinelearning #AI

1

It is fascinating to witness the battle between the AQLM (Additive Quantization of Language Models) and QuIP# (Quantization with Incoherence Processing) teams in the 2-3-bit #quantization art.

1

I get very concerned when I see ML scientists/engineers being wowed by basic math and make statements like “I did not realize until today that ReLU activation function is not differentiable at x = 0”. If you did not know this, then you certainly are blissfully unaware of how the non-differentiability issue is handled in back-prop, despite using ReLU (and Leaky ReLU) on a regular basis. “Attention is not all you need - you need some mathematical fundamentals as well to be a good ML scientist” :-)

1,459

117 Comments