Class HNSWConfig

vector/beir/config.py:36–77

HNSW index configuration.

@dataclass
class HNSWConfig:
    """HNSW index configuration."""
    metric: str = os.getenv("HNSW_METRIC", "euclidean")
    max_levels: int = int(os.getenv("HNSW_MAX_LEVELS", "3"))
    ef_search: int = int(os.getenv("HNSW_EF_SEARCH", "40"))
    ef_construction: int = int(os.getenv("HNSW_EF_CONSTRUCTION", "100"))
    # Index type allows switching between standard hnsw and experimental variants
    # such as the new "partionedhnsw" (sic) index.
    index_type: str = os.getenv("HNSW_INDEX_TYPE", "hnsw")
    # Optional number of clusters for partitioned HNSW experimental variants. Interpreted as a
    # string in the index definition to match existing HNSW parameters. Only set
    # when HNSW_NUM_CLUSTERS is provided; otherwise default to 0 (disabled).
    num_clusters: int = int(os.getenv("HNSW_NUM_CLUSTERS")) if os.getenv("HNSW_NUM_CLUSTERS") is not None else 0
    # Optional vector dimension; when using partionedhnsw (sic) this should match the
    # length of the stored embeddings. Experimental, this is not yet released.
    vector_dim: int = 0
    
    def to_index_string(self) -> str:
        """Generate index string for Dgraph schema."""
        # Base index name (e.g., "hnsw" or "partionedhnsw")
        index_name = self.index_type

        # Common parameters shared by both standard and partitioned HNSW
        params = [
            f'metric:"{self.metric}"',
            f'maxLevels:"{self.max_levels}"',
            f'efSearch:"{self.ef_search}"',
            f'efConstruction:"{self.ef_construction}"',
        ]

        # Partitioned HNSW adds numClusters; we include it whenever explicitly
        # configured, regardless of the index name, to keep behavior simple.
        if self.num_clusters > 0:
            params.append(f'numClusters:"{self.num_clusters}"')

        # Some partitioned HNSW variants require the vector dimension to be
        # specified explicitly in the index configuration.
        if self.index_type == "partionedhnsw" and self.vector_dim > 0:
            params.append(f'vectorDimension:"{self.vector_dim}"')

        joined = ",".join(params)
        return f"{index_name}({joined})"


@dataclass