multi_retriever_usage.md

Multi-Retriever Usage

Configuration

Using multiple retrievers simultaneously requires configuration in the multi_retrieve section of the config file. Enable use_multi_retriever and configure it as follows:

use_multi_retriever: True # whether to use multi retrievers
multi_retriever_setting:
  merge_method: "concat" # support 'concat', 'rrf', 'rerank'
  topk: 5 # final remaining documents, only used in 'rrf' and 'rerank' merge
  rerank_model_name: ~
  rerank_model_path: ~
  retriever_list:
    - retrieval_method: "e5"
      retrieval_topk: 5
      index_path: ~
      retrieval_model_path: ~
    - retrieval_method: "bm25"
      retrieval_topk: 5
      index_path: ~
      retrieval_model_path: ~

The retriever_list can include multiple retrievers, each with its own corresponding corpus and index. The configuration for each retriever follows the same settings as a single retriever.

Currently, three aggregation methods are supported:

1. concat: Directly concatenate the results from multiple retrievers.
2. rrf: Aggregate the results from retrievers using the RRF (Reciprocal Rank Fusion) algorithm.
3. rerank: Use a reranker to rerank all the results from the retrievers. This requires configuring the reranker.

For both rrf and rerank, only the top k results will be retained.

Quick Usage

Below is a quick example of how to use the multi-retriever feature:

from flashrag.utils import get_retriever
from flashrag.config import Config

config_dict = {
    "gpu_id": "1",
    "use_multi_retriever": True,
    "multi_retriever_setting": {
        "merge_method": "rerank",
        "topk": 5,
        'rerank_model_name': 'bge-reranker',
        'rerank_model_path': 'bge-reranker-m3',
        "retriever_list": [
            {
                "retrieval_method": "bm25",
                "corpus_path": "general_knowledge.jsonl",
                "index_path": "indexes/general_knowledge/bm25",
                "retrieval_topk": 3,
                "bm25_backend": "pyserini",
            },
            {
                "retrieval_method": "e5",
                "corpus_path": "general_knowledge.jsonl",
                "index_path": "indexes/general_knowledge/e5_Flat.index",
                "retrieval_topk": 1,
            },
        ],
    },
}
config = Config("my_config.yaml", config_dict=config_dict)
retriever = get_retriever(config)
query_list = ['who is the president of USA?']

output, score = retriever.batch_search(query_list, return_score=True)
for item,s in zip(output, score):
    print(item, s)
    print("----")