Implementation of Rerank for RankVicuna

[yilinjz]

Rerank can take either query + docs or query + hits; examples can be found in the newly added demo/ folder.

The following fields of RankVicuna are given default values:

• context_size: int = 4096,
• prompt_mode: PromptMode = PromptMode.RANK_GPT,
• device: str = "cuda",
• num_gpus: int = 1,

Removed two fields from RankLLM (top_k_candidates and dataset); RankVicuna currently does not use those two flags, and since RankVicuna inherited from RankLLM (and RankLLM does not use them either) I removed them from RankLLM as well.

[sahel-sh]

Rank Vicuna and Rank_LLM already have a rerank that gets a dataset, retrieves its top k candidates and then uses the provided model to rerank them. Please look at run_rank_llm.py to see how the current functionally works. If you want the added functionality for reranking the provided documents, or hits, it should not replace the current functionality. But added as extra.

[ronakice]

I feel it is fine to disentangle these? Rerank should ideally only work on reranking and not entangled with retriever? We should reorganise And have say another retrieve_and_rerank method that mirrors what you have but in a different module.

[yilinjz]

@sahel-sh @ronakice Updated the PR with changes we discussed last week.

A few questions:

1. write_rerank_result() method is now moved to the new Reranker class; do we still want to include "dataset" and "top_k_candidates" as part of the output file name?
2. I wrapped the 3 cases of retrieving into one retrieve() method in the new Retriever class, do you think this way is okay?

[yilinjz]

Accidentally closed this when resolving a conflict... made another PR here https://github.com/castorini/rank_llm/pull/24