Seonglae ChoRetrieval-Augmented Generation
RAG is not Memory, Search or Memory is better than naive RAG. Better retrieval performance ≠ Better performance
Challenges
- retrieval latency
- system complexity
RAG Notion
RAG Usages
Beyond dot product
Stanford CS25: V3 I Retrieval Augmented Language Models
December 5, 2023
Douwe Kiela, Contextual AI
Language models have led to amazing progress, but they also have important shortcomings. One solution for many of these shortcomings is retrieval augmentation. I will introduce the topic, survey recent literature on retrieval augmented language models and finish with some of the main open questions.
More about the course can be found here: https://web.stanford.edu/class/cs25/
View the entire CS25 Transformers United playlist: https://www.youtube.com/playlist?list=PLoROMvodv4rNiJRchCzutFw5ItR_Z27CM
https://www.youtube.com/watch?v=mE7IDf2SmJg&list=PLoROMvodv4rNiJRchCzutFw5ItR_Z27CM&index=25
Intro of Retrieval Augmented Generation (RAG) and application demos
Introduction of Retrieval Augmented Generation, Jupyter Notebook three demos of Basic RAG, Sentence-window retrieval, Auto-merging…
https://medium.com/@henryhengluo/intro-of-retrieval-augmented-generation-rag-and-application-demos-c1d9239ababf
NVIDIA Research: RAG with Long Context LLMs
This blog post dives into NVIDIA’s recent study comparing retrieval-augmentation with and without long-context LLMs.
https://blog.llamaindex.ai/nvidia-research-rag-with-long-context-llms-7d94d40090c4
KBQA, Knowledge Graph
ML Blog - Improve ChatGPT with Knowledge Graphs
Leveraging knowledge graphs for LLMs using LangChain
https://mlabonne.github.io/blog/posts/Article_Improve_ChatGPT_with_Knowledge_Graphs.html
llama-recipes/recipes/use_cases/agents/langchain at main · meta-llama/llama-recipes
Scripts for fine-tuning Meta Llama3 with composable FSDP & PEFT methods to cover single/multi-node GPUs. Supports default & custom datasets for applications such as summarization an...
https://github.com/meta-llama/llama-recipes/tree/main/recipes/use_cases/agents/langchain
Base model is better at retrieval than Instruction Tuning model
arxiv.org
https://arxiv.org/pdf/2406.14972
Passages retrieved through RAG search often have little relation to each other, resulting in attention forming a block-diagonal structure. Therefore, the argument is that feeding all tokens directly into the decoder is computationally wasteful.
arxiv.org
https://arxiv.org/pdf/2509.01092
Beyond RAG: Terminal is all you need
Beyond Semantic Similarity: Rethinking Retrieval for Agentic...
Modern retrieval systems, whether lexical or semantic, expose a corpus through a fixed similarity interface that compresses access into a single top-k retrieval step before reasoning. This...
https://arxiv.org/abs/2605.05242
Terminal Is All You Need: Design Properties for Human-AI Agent Collaboration
While research on AI agents focuses on enabling them to operate graphical user interfaces, the most effective and widely adopted agent tools in practice are terminal-based. We argue that this convergence is not coincidental. It reflects three design properties central to effective human-AI-UI collaboration: representational compatibility between agent and interface, transparency of agent actions within the interaction medium, and low barriers to entry for human participants. We ground each property in established HCI theory, show how terminal-based tools satisfy them by default, and argue that any modality, including graphical and spatial interfaces, must be deliberately engineered to achieve them. Rather than a legacy artifact, the terminal serves as a design exemplar whose properties any agent-facing modality must replicate.
https://arxiv.org/html/2603.10664v1