Provence

[nadiinchi]

• add Provence context_processor and configs
• add baseline context pruners context_processors and configs
• add context compression calculation & saving in a separate metric file in an exp folder
• little refactor of modules/rag.py to better accommodate context processing
• add possibility to provide oracle docs in the query dataset
• add RGB dataset (with oracle docs in query dataset)

[DRRV]

just be to usre: here Provence does not do reranking right?

[nadiinchi]

just be to usre: here Provence does not do reranking right?

it does. If reorder=True in provence config, provence reorders documents in the context processing list. But it is incorporated into context processing, so you would set retriever=<smth> context_processor=provence/provence_reranking_0.1 and skip reranker

[maxime-louis]

Two consecutive line returns: only one is allowed within functions

[maxime-louis]

I would use 'prob.item()' to make it a float and not a torch tensor

[maxime-louis]

just be to usre: here Provence does not do reranking right?

it does. If reorder=True in provence config, provence reorders documents in the context processing list. But it is incorporated into context processing, so you would set retriever=<smth> context_processor=provence/provence_reranking_0.1 and skip reranker

Can you put an assert for that ?

[sclincha]

1. Do we want to update the requirements file for LLMLingua? Or Simply Try Catch, please install?
2. It could be good to have a simple README for context processor explaining the list of models and basic usage such as:
   CONFIG=myconfig python bergen.py retriever=splade-v3 reranker=debertav3 +context_processor=provence/provence_standalone_0.1 dataset=$1 generator=vllm_llama-2-7b-chat
   What do you think? Where should we put that README?. The run_exp file is a bit hard to read for a first usage ...