This video explains the details behind Active Retrieval Augmented Generation from Jiang et al! This is a super exciting innovation on how exactly we construct retrieval-augmented generation pipelines that hand off from Weaviate to Large Language Models! Please see the links below for more information about particular parts of the paper analysis! Please leave a like and subscribe if you want to see more content like this, thank you so much for watching!
Links:
Active Retrieval Augmented Generation - arxiv.org/pdf/2305.06983.pdf
OpenAI Playground - platform.openai.com/playground
2WikiMultiHopQA - aclanthology.org/2020.coling-...
ASQA - aclanthology.org/2022.emnlp-m...
StrategyQA - arxiv.org/pdf/2101.02235.pdf
WikiAsp - (sorry the url link was too long, this is from Hayashi et al.)
UniEval - aclanthology.org/2022.emnlp-m...
Chapters
0:00 Welcome!
0:25 FLARE in 2 Minutes!
3:12 Retrieval-Augmented Generation
4:44 The Intuition
7:17 Algorithm Details
8:23 When to Query
9:50 Log Prob from OpenAI!!!!
11:58 How to Query
14:51 Baseline Approaches
17:00 Self-Ask Prompting
18:18 FLARE[indirect] (Search Action Ablation)
25:35 Evaluation Datasets
28:05 2WikiMultihopQA
29:55 StrategyQA
31:22 ASQA
32:10 WikiAsp
32:48 Summary of Evaluation Datasets
34:14 Evaluation Metrics
36:45 UniEval
37:54 Results!
41:45 Personal Reflections and Ideas