Data Management:RetrieveKnowledgeBase

The ID of the knowledge base.

The query text.

The number of top-ranked results to return.

A filter for the data, specified as a SQL WHERE clause.

The distance metric for retrieval. If unspecified, this defaults to the metric configured for the knowledge base. Only set this parameter if you have specific requirements.

Valid values:

• l2: Euclidean distance.

• ip: Inner product.

• cosine: Cosine similarity.

The hybrid search algorithm. If this parameter is not set, the system directly compares and ranks the scores from the dense vector and full-text searches.

Valid values:

• RRF: Reciprocal Rank Fusion. This method uses a parameter k to control the fusion effect. For more information, see the HybridSearchArgs configuration.

• Weight: Weighted ranking. This method applies weights to the vector and full-text search scores before ranking. For more information, see the HybridSearchArgs configuration.

• Cascaded: Performs a full-text search first, followed by a vector search on the results of the full-text search.

Parameters for the specified HybridSearch algorithm. Both RRF and Weight are supported. You can use the HybridPathsSetting object to specify the retrieval paths: dense vector (dense), sparse vector (sparse), and full-text search (fulltext). If this object is not provided, the default retrieval paths are dense and fulltext.

• RRF: Specifies the constant k in the scoring formula 1/(k+rank_i). The value of k must be an integer greater than 1. The format is as follows:

{
  "HybridPathsSetting": {
    "paths": "dense,fulltext"
  },
  "RRF": {
    "k": 60
  }
}

• Weight:
  • Two-path recall (do not specify HybridPathsSetting; specify only alpha):

    • The score is calculated using the formula: alpha * dense_score + (1-alpha) * fulltext_score. The alpha parameter balances the scores from the dense vector and full-text searches. Its value must be in the range [0, 1], where 0 relies solely on full-text search, and 1 relies solely on dense vector search.

{ 
   "Weight": {
    "alpha": 0.5
   }
}

• Three-path recall:

  • The score is calculated using the formula: normalized_dense * dense_score + normalized_sparse * sparse_score + normalized_fulltext * fulltext_score. The dense, sparse, and fulltext parameters are the weights for the dense vector, sparse vector, and full-text searches, respectively. Their values must be 0 or greater. The system automatically normalizes the weights to sum to 1 (for example, normalized_x = x / (dense + sparse + fulltext)).

{
  "HybridPathsSetting": {
     "paths": "dense,sparse,fulltext"
   },
  "Weight": {
    "dense": 0.5,
    "sparse": 0.3,
    "fulltext": 0.2
  }
}

Specifies whether to include the vector in the results. The default value is false.

The metadata fields to return, separated by commas. By default, no metadata fields are returned.

The factor used to rerank vector search results. The value must be in the range (1, 5].

The recall window. If specified, this parameter expands the context of the retrieved results. The format is a two-element array [A, B], where -10 <= A <= 0 and 0 <= B <= 10.

The field to use for sorting the results. By default, this parameter is empty.

The field must be a metadata field or a default table field, such as id. Supported formats include:

You can specify a single field (for example, chunk_id), multiple comma-separated fields (for example, block_id, chunk_id), or fields with descending order (for example, block_id DESC, chunk_id DESC).

The offset for pagination.