​
Configuration Parameters

  • retriever_path (str): Path or name of the embedding model
  • corpus_path (str): Path to the corpus file (JSONL format)
  • embedding_path (str): Path to the .npy file storing generated embedding vectors
  • index_path (str): Path to the .index file storing the Faiss index
  • infinity_kwargs (dict): Model configuration passed to infinity_emb.EngineArgs
    • bettertransformer (bool): Whether to enable BetterTransformer acceleration (default false)
    • pooling_method (str): Vector pooling method (auto, cls, mean)
    • device (str): Device to load the model on (e.g., cuda, cpu)
    • batch_size (int): Batch size when embedding in batches
  • cuda_devices (str): Visible GPU settings, e.g., "0,1" means using GPU 0 and GPU 1
  • query_instruction (str): Prefix prompt for queries
  • faiss_use_gpu (bool): Whether to enable GPU Faiss (fallback to CPU if failed)
  • top_k (int): Number of retrieved documents to return
  • overwrite (bool): Whether to overwrite if the target exists
  • retriever_url (str): Service address (can be host:port or full http://host:port, default port 8080 if not specified)
  • index_chunk_size (int): Batch size for add_with_ids in chunks

​
API Description

​
retriever_init

​
Function

Initialize the retriever model based on the Infinity Embedding library and load the corpus file; if index_path is provided and exists, load the Faiss index (GPU/CPU).

​
Input Parameters

  • retriever_path (str): Infinity model path or name (passed to EngineArgs)
  • corpus_path (str): Corpus file (JSONL), each line must contain the key "contents"
  • index_path (Optional[str]): Existing .index file path of the built index, load if exists
  • faiss_use_gpu (bool): Whether to enable GPU Faiss (fallback to CPU if failed)
  • infinity_kwargs (Optional[Dict[str, Any]]): Other parameters passed to EngineArgs (e.g., dtype, batch_size, etc.)
  • cuda_devices (Optional[str]): Visible GPUs, e.g., "0,1"; will be set to CUDA_VISIBLE_DEVICES

​
Return Parameters

  • None

​
retriever_embed

​
Function

Encode the corpus using the retriever model into embeddings and save as .npy.

​
Input Parameters

  • embedding_path (Optional[str]): Output .npy file path; if not provided, defaults to <project_root>/output/embedding/embedding.npy
  • overwrite (bool): Whether to overwrite if the target exists

​
Return Parameters

  • None

​
retriever_index

​
Function

Build Faiss Index (IndexIDMap2(FlatIP)) based on .npy embeddings, supports chunked insertion and GPU construction.

​
Input Parameters

  • embedding_path (str): File storing the encoded embeddings, .npy file
  • index_path (Optional[str]): Output .index file path; defaults to <project_root>/output/index/index.index if not provided
  • overwrite (bool): Whether to overwrite if exists
  • index_chunk_size (int): Batch size for add_with_ids in chunks

​
Return Parameters

  • None

​
Function

Perform vector retrieval on the Faiss index.

​
Input Parameters

  • query_list (List[str]): List of queries or a single string (automatically wrapped into a list)
  • top_k (int): Number of results to return
  • query_instruction (str): Prefix prompt added to each query (instruction-based query)
  • use_openai (bool): Whether to use OpenAI to generate query vectors; otherwise use the Infinity model by default

​
Return Parameters

  • ret_psg (Dict[str, List[List[str]]]): Retrieved passages

​
retriever_deploy_service

​
Function

Start a lightweight Flask service to deploy the retriever model.

​
Input Parameters

  • retriever_url (str): Service address (can be host:port or full http://host:port, default port 8080 if not specified)

​
Return Parameters

  • None

​
Function

Call the remote /search endpoint of the retriever_deploy_service as a client and return remote retrieval results.

​
Input Parameters

  • retriever_url (str): Remote base address
  • query_list (List[str]): List of queries or a single string (automatically wrapped into a list)
  • top_k (int): Number of results to return
  • query_instruction (str): Prefix prompt added to each query (instruction-based query)

​
Return Parameters

  • ret_psg (Dict[str, List[List[str]]]): Retrieved passages