I've noticed more companies stepping back from building RAG systems. Reconstructing a decades-old discipline like information retrieval isn't easy, even with advanced LLMs. Additionally, evaluating and monitoring RAG pipelines remains challenging and unresolved. An often overlooked point: you don't necessarily need vector search to build a #RAG. Traditional methods like #BM25 can be more effective, consistently performant, and much cheaper. While retrieving certain facts from a collection of documents is fairly straightforward, the most ambitious promise of RAG—answering open-ended questions grounded in a large body of multimodal content—remains unfulfilled. This is challenging because relevant documents might not be semantically close to the question in the embedded space. A potential solution could be using LLMs' internal knowledge to "seed" various searches, combine the results, and then use LLM reasoning to sift through a large context to refine and answer the question. Is anyone exploring this approach?
Mehdi Jamei’s Post
More Relevant Posts
-
◉ Senior Solution Architect & Innovation at Expleo Group | Business Development, Value-led Operation Transformation, Consulting & Management to deliver value to all stakeholders in the Aeronautics Industry
A nice link list about #LLM subject that goes straight to the point. It gathers interesting resources to understand the different technologies and solutions.
-
Here is the third and final part of the lecture on Unstructured Navier-Stokes solver development. https://lnkd.in/eM96ZjiA #CFD
SIMPLE Algorithm for an Unstructured Mesh--Part 3 of 3.
https://www.youtube.com/
-
RAG systems improve LLMs by adding memory and context to provide more accurate and applicable responses in complex settings. Naive RAG, indexing documents for response generation, faced issues like low precision and outdated data, setting the stage for future advancements. Advanced RAG methods enhance pre-retrieval, retrieval, and post-retrieval, introducing precise strategies like Adaptive Retrieval and Hypothetical Document Embeddings for a more meticulous process. The Modular RAG introduces a flexible, customizable setup with modules like Search, Memory, Fusion, and more, enhancing adaptability and efficiency in LLMs for more relevant responses. Let’s understand how RAG systems evolve in LLM applications through the transition from Naive to Advanced RAGs: https://lnkd.in/gVSpWQmz
-
-
Founder & CEO | Seasoned data and AI/ML executive in healthcare, retail, and hospitality
OpenAI shares their best practices in RAG (Retrieval-Augmented Generation) technology for customers with large document collections, such as 100,000 documents, who need models to perform knowledge retrieval based on these documents. 1. Initially, directly embedding PDF and docx files resulted in an accuracy rate of 45%. 2. After 20 iterations of tuning to fix minor bugs, the accuracy rate improved to 65%. 3. By optimizing with rules, such as first determining the field the question pertains to before answering, the accuracy rate increased to 85%. 4. Recognizing that some data was structured (e.g., tables), a customized extraction solution was developed, further boosting the accuracy to 98%. https://lnkd.in/egnaBwrm
A Survey of Techniques for Maximizing LLM Performance
https://www.youtube.com/
-
In response to many requests over the past couple of years, here is Part 1 of the lecture series on Unstructured Navier-Stokes solver. I will post the next two parts shortly (apologies in advance). #CFD
SIMPLE Algorithm for an Unstructured Mesh--Part 1 of 3.
https://www.youtube.com/
-
Here are a list of things I've found interesting to read about LLMs, over the last few months - https://lnkd.in/ek6VrU8m
-
📄 The Prompt Report: A Systematic Survey of Prompting Techniques A very nice weekend find, covering a structured understanding of prompts and a taxonomy of 58 text-only prompting techniques, including zero-shot, few-shot, thought generation, ensembling, self-criticism and decomposition. The linked paper/PDF gives a very detailed overview of all common prompt techniques and evaluates their applicability and quality (Few-shot Chain-of-Thought performs the best among the techniques they benchmarked). The insights into prompt engineering, answer engineering, security, agents and alignment are detailed and should be in every "GenAI Toolbox". Enjoy! https://lnkd.in/edzVD4H9
-
Executive Leader | Data Polymath | Payments | AI Strategy | Engineering | Product | Innovation | People
2wWould using LLM for problems which cannot be solved by traditional means would be a good start. Also LLM needs to be an enabler as you pointed out instead of the end state