constructor

Configuration options for setting up the retriever:

• vectorStore (required): The VectorStore instance implementing VectorStoreInterface that will be used to store and retrieve document embeddings. This is the core component of the retriever, enabling vector-based similarity and MMR searches.

• k (optional): Specifies the number of documents to retrieve per search query. If not provided, defaults to 4. This count determines the number of most relevant documents returned for each search operation, balancing performance with comprehensiveness.

• searchType (optional): Defines the search approach used by the retriever, allowing for flexibility between two methods:

  • "similarity" (default): A similarity-based search, retrieving documents with high vector similarity to the query. This type prioritizes relevance and is often used when diversity among results is less critical.
  • "mmr": Maximal Marginal Relevance search, which combines relevance with diversity. MMR is useful for scenarios where varied content is essential, as it selects results that both match the query and introduce content diversity.

• filter (optional): A filter of type FilterType, defined by the vector store, that allows for refined and targeted search results. This filter applies specified conditions to limit which documents are eligible for retrieval, offering control over the scope of results.

• searchKwargs (optional, applicable only if searchType is "mmr"): Additional settings for configuring MMR-specific behavior. These parameters allow further tuning of the MMR search process:

  • fetchK: The initial number of documents fetched from the vector store before the MMR algorithm is applied. Fetching a larger set enables the algorithm to select a more diverse subset of documents.
  • lambda: A parameter controlling the relevance-diversity balance, where 0 emphasizes diversity and 1 prioritizes relevance. Intermediate values provide a blend of the two, allowing customization based on the importance of content variety relative to query relevance.