thoughts

[Jason Liu]

By what dimensions do they get more complicated? I thought we were just chatting with PDFs.

summarization
processing tables
processing images
query against the metadata vs the content
query against time
leverage structure at query expansion time
leveraging structure at q/a time
citations to ground models
citations to finetune embedding models
using rerankers and hybred search
synthetic data generation for early evaluations
understanding which metrics to use
communicating itermediate steps to to the user to reduce their sensititivty to latency
monitoring in production
how to slice metrics as you onboard new users.
how to topic model inbound questions to determine where do invest more processing

[Grzegorz Sajko]

I think that beyond level 5 every next level is paralel

flowchart TB

lv1 --> lv2 --> lv3 --> lv4 --> lv5

lv5 --> lv_6 & lv_7 & lv_8 & lv_x4 & lv_x5

Loading

Also you have to be mindful how deep you want to go, since I think you'll just would start doing Information Retrieval (I know it's a R in RAG, but I think for some topics, I would be useful to just mention them, not go deep into)

I like this blogpost that overviews many RAG techniques
https://pub.towardsai.net/advanced-rag-techniques-an-illustrated-overview-04d193d8fec6
maybe you could find it useful