ViscondeRunPipeline

Overview

The ViscondeRunPipeline class is implementation of Visconde pipeline. You can check out Visconde paper at here.

Visconde pipeline perform three task: decompose, retrieve, and aggregate. It uses Query Decomposition for answering multi-hop questions. So, it is effective to answer real-world questions that need to check out multiple passages.

Usage

Initialize

To create an instance of ViscondeRunPipeline, you need to provide an instance of a Retrieval class, and llm module to generate answer. Optionally, you can specify the instance of query decomposition module, a custom prompt, and other options for retrieval and use passage count for generation. FYI, you can't use chat model for this pipeline. It has a default prompt for strategyQA style multi-hop questions. You need to change prompt using PromptTemplate if you want to use another few-shot prompts.

from RAGchain.pipeline import ViscondeRunPipeline
from RAGchain.retrieval import BM25Retrieval
from langchain.llms.openai import OpenAI

retrieval = BM25Retrieval(save_path="path/to/your/bm25/save_path")
pipeline = ViscondeRunPipeline(retrieval, OpenAI(model_name="babbage-002"))

Ask

You can ask a question to the LLM model and get an answer using invoke method. Also, you can use another LCEL's method like stream or batch as well.

question = "Is reranker and retriever have same role?"
answer = pipeline.run.invoke({"question": question})
print(answer)

If you want to get used passages or relevance scores of retrieved passages, you can use get_passages_and_run method.

answers, passages, scores = pipeline.get_passages_and_run([question])