#52 Here’s How We Are Improving Search in RAG Systems

Which Search Technique Should You Use: Metadata Filtering, GraphRAG, Routers, or a Simple Embedding Search

Good morning, AI enthusiasts! This week, we start our discussion with RAG (if you feel we talk too much about it, you don’t know how amazing RAG is); we also have a practical tutorial exploring the potential of Multimodal RAG and an essential discussion on LLM vulnerabilities and how to mitigate them. We also have an interesting product from the community, fun collaboration opportunities, and, of course, some more exciting resources and discussions. Enjoy the read! 

What’s AI Weekly

This week in What’s AI, we are exploring how to improve how you find information in your RAG applications. We’ll cover everything from old-school keyword searches to more recent methods such as GraphRAG. By the end, you’ll know exactly how to make your data work for you. Read the complete article here or watch the video on YouTube! 

— Louis-François Bouchard, Towards AI Co-founder & Head of Community

Learn AI Together Community section!

Featured Community post from the Discord

Jjjaaackkk built a modern Jupyter client for Mac. It manages Python versions, dependencies, and virtual environments (by wrapping uv). It offers faster startup times than VS Code and JupyterLab, generates code inline with context-aware prompt cells, automatically detects existing kernels and conda environments, and more. Check it out here and support a fellow community member. Share your questions and feedback in the thread! 

AI poll of the week!

It seems our community has a strong entrepreneurial interest. If you could achieve your top AI career goal in 2024, what would be your first big milestone? Share your plans in the thread, and we would love to help you figure out how to get there quickly. 

Collaboration Opportunities 

The Learn AI Together Discord community is flooding with collaboration opportunities. If you are excited to dive into applied AI, want a study partner, or even want to find a partner for your passion project, join the collaboration channel! Keep an eye on this section, too—we share cool opportunities every week! 

1. Dykyi_vladk is working on reimplementing and enhancing the PaLM model. If you’re also interested in NLP, reach out in the thread! 

2. Abdurrahman01234 wants to build practical projects for his portfolio. He is looking for someone to brainstorm ideas with. If this sounds exciting, connect with him in the thread! 

3. Tonystank1250 is looking for a collaborator for a personal project. If you have some time to spare, contact him in the thread! 

Meme of the week!

Meme shared by ghost_in_the_machine

TAI Curated section

Article of the week

Generalized Additive Nonlinear Models for Time Series Forecasting By Shenggang Li

This article presents a novel time series forecasting model, GAM-ARMA, which integrates bounded nonlinear functions into the traditional ARMA framework. The core innovation is replacing linear autoregressive terms with a bounded nonlinear function, improving stability and interpretability. This approach addresses the limitations of ARMA's linearity assumption and the data requirements of neural networks, particularly beneficial with sparse data. Experiments using the Electric Production dataset demonstrate that GAM-ARMA outperforms traditional ARMA models across various forecast horizons and training window sizes, as measured by RMSE and MAPE, showcasing its enhanced accuracy and robustness. It further details the model's mathematical framework, optimization procedure, and implementation, highlighting its advantages and potential extensions for multivariate time series forecasting and uncertainty quantification. 

Our must-read articles

1. Compression is Generalisation, Generalisation is Intelligence - Unsupervised Learning in Large Language Models By Alex Punnen

This article explores unsupervised learning in large language models (LLMs), arguing that their ability to generalize stems from the compression of information during training. While LLMs predict the next word in a sequence, their success isn't solely probabilistic; they build a world model, enabling reasoning and understanding. The author contrasts this with supervised learning's limitations, highlighting the scale of model parameters and training data as crucial factors for emergent capabilities. Using Kolmogorov complexity, it illustrates how compression reflects the model's understanding. It also discusses the empirical evidence of unsupervised learning's success in GPT-2 and GPT-3, demonstrating their improved performance across various NLP tasks compared to supervised approaches.

2. Llama-OCR + Multimodal RAG + Local LLM Python Project: Easy AI/Chat for your Docs By Gao Dalie (高達烈)

This article presents a tutorial on building a local chatbot using Llama-OCR, Multimodal RAG, and a local LLM. It addresses the limitations of traditional OCR in handling complex document layouts. It uses ColPali, a multimodal retrieval system using visual language models to directly encode images for efficient text extraction from PDFs without needing OCR. The process involves indexing a document with ColQwen2, querying it, retrieving the relevant page, and then using the Llama 3.2 Vision model (via Ollama) to process the image and answer the query. It provides a step-by-step guide with code examples demonstrating the implementation.

3. LLM Agent Jailbreaking and Defense — 101 By Mohit Sewak, Ph.D.

This article explores the vulnerabilities of large language models (LLMs), such as jailbreaking. It details various methods, including prompt engineering, data poisoning, and exploiting agentic workflows, used by hackers to manipulate LLMs into undesirable behavior. Conversely, the article also outlines defensive strategies employed to mitigate these risks, such as robust training, safety guidelines, human oversight, sandboxing, and input/content filtering. Finally, it discusses ongoing research focusing on proactive defenses, including enhanced explainability and adaptive measures to combat evolving threats like multi-agent jailbreaking and adaptive attacks.

If you are interested in publishing with Towards AI, check our guidelines and sign up. We will publish your work to our network if it meets our editorial policies and standards.