Uploaded December 2024 | Updated September 2026, 3 weeks ago
Retrieval-Augmented Generation (RAG) empowers large language models (LLMs) by integrating external knowledge, boosting their ability to generate informed and accurate outputs. However, as #llms process longer inputs and retrieve more #data, challenges like declining quality from "hard negatives" arise.
Bowen Jin from the University of Illinois Urbana-Champaign dives into why this happens and how it can be fixed. He shares how retrieval reordering can be a simple, effective optimization method, alongside advanced fine-tuning approaches that boost performance. Check out our talk to explore new strategies to improve AI by refining data sources, retriever design, and training processes.
Timestamps:
0:00 Introduction
0:43 What is RAG, and how does it support various applications?
1:34 LLMs are able to support longer context
2:22 RAG or Long-context LLM?
4:37 Long-context LLM in RAG
5:10 The effect of retrieved context size on RAG performance
8:21 The interplay of retrieval quality and LLM capabilities
13:45 The importance of hard negatives for long-context LLM evaluation
19:35 Simple and effective training-free RAG improvement
22:42 Improving Robustness for RAG via Data-Augmented Fine-Tuning
26:45 Case study
27:45 Takeaways
29:41 Q&A
#rag #llms #ai #retrieval #artificialintelligence #longcontextAI #aioptimization #datascience #machinelearning #deeplearning #externalknowledge #finetuning #airesearch #dataselection #reordering #optimization #retrievalaugmentedgeneration #RAG #technology #tech #techtalk #techtalks #aitalks #aitalk #science #python #pythonprogramming #ai #programming
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
Retrieval-Augmented Generation (RAG) empowers large language models (LLMs) by integrating external knowledge, boosting their ability to generate informed and accurate outputs. However, as #llms process longer inputs and retrieve more #data, challenges like declining quality from "hard negatives" arise.
Bowen Jin from the University of Illinois Urbana-Champaign dives into why this happens and how it can be fixed. He shares how retrieval reordering can be a simple, effective optimization method, alongside advanced fine-tuning approaches that boost performance. Check out our talk to explore new strategies to improve AI by refining data sources, retriever design, and training processes.
Timestamps:
0:00 Introduction
0:43 What is RAG, and how does it support various applications?
1:34 LLMs are able to support longer context
2:22 RAG or Long-context LLM?
4:37 Long-context LLM in RAG
5:10 The effect of retrieved context size on RAG performance
8:21 The interplay of retrieval quality and LLM capabilities
13:45 The importance of hard negatives for long-context LLM evaluation
19:35 Simple and effective training-free RAG improvement
22:42 Improving Robustness for RAG via Data-Augmented Fine-Tuning
26:45 Case study
27:45 Takeaways
29:41 Q&A
#rag #llms #ai #retrieval #artificialintelligence #longcontextAI #aioptimization #datascience #machinelearning #deeplearning #externalknowledge #finetuning #airesearch #dataselection #reordering #optimization #retrievalaugmentedgeneration #RAG #technology #tech #techtalk #techtalks #aitalks #aitalk #science #python #pythonprogramming #ai #programming
Social Links:
Newsletter: buzzrobot.substack.com
X: https://x.com/sopharicks
Slack: join.slack.com/t/buzzrobot/shared_invite/zt-2s067rv7n-guPIMGe62rbp9ncxdnOUfQ
