What causes "Unsupported metric_type: {distance}" in mem0ai/mem0?

Pass the exact enum member name in uppercase, e.g. distance='COSINE' (check pymochow's MetricType for the accepted names such as COSINE/L2/IP).

mem0ai/mem0 · error · ValueError

Unsupported metric_type: {distance}

Error message

Unsupported metric_type: {distance}

What it means

BaiduDB.create_col raises ValueError when the distance string does not exactly match any pymochow MetricType enum member NAME. The loop compares the raw string against MetricType.__members__ keys (e.g. 'COSINE', 'L2', 'IP'), so it is case-sensitive: 'cosine' fails while 'COSINE' passes, and unsupported metrics like 'euclidean' fail outright.

Source

Thrown at mem0/vector_stores/baidu.py:133

            vector_size (int): Dimension of the vector.
            distance (str): Metric type for similarity search.
        """
        # Check if table already exists
        try:
            tables = self._database.list_table()
            table_exists = any(table.table_name == name for table in tables)
            if table_exists:
                logger.info(f"Table {name} already exists. Skipping creation.")
                self._table = self._database.describe_table(name)
                return

            # Convert distance string to MetricType enum
            metric_type = None
            for k, v in MetricType.__members__.items():
                if k == distance:
                    metric_type = v
            if metric_type is None:
                raise ValueError(f"Unsupported metric_type: {distance}")

            # Define table schema
            fields = [
                Field(
                    "id", FieldType.STRING, primary_key=True, partition_key=True, auto_increment=False, not_null=True
                ),
                Field("vector", FieldType.FLOAT_VECTOR, dimension=vector_size),
                Field("metadata", FieldType.JSON),
            ]

            # Create vector index
            indexes = [
                VectorIndex(
                    index_name="vector_idx",
                    index_type=IndexType.HNSW,
                    field="vector",
                    metric_type=metric_type,
                    params=HNSWParams(m=16, efconstruction=200),

View on GitHub (pinned to 001c235229)

Solutions

Pass the exact enum member name in uppercase, e.g. distance='COSINE' (check pymochow's MetricType for the accepted names such as COSINE/L2/IP).
If the string comes from shared config, uppercase/normalize it before it reaches BaiduDB: distance.strip().upper().
Verify available names at runtime: python -c "from pymochow.model.table import MetricType; print(list(MetricType.__members__))".

Example fix

# before
BaiduDB(..., distance_metric or config with distance="cosine")  # ValueError

# after
config = {"distance": "COSINE", ...}  # exact MetricType member name

Defensive patterns

Strategy: validation

Validate before calling

from pymochow.model.table import MetricType

VALID_METRICS = set(MetricType.__members__)

def normalize_distance(distance: str) -> str:
    d = distance.strip().upper()
    if d not in VALID_METRICS:
        raise ValueError(f"distance must be one of {sorted(VALID_METRICS)}, got {distance!r}")
    return d

distance = normalize_distance(cfg["distance"])  # before creating BaiduDB

Type guard

def is_valid_baidu_metric(distance: str) -> bool:
    return isinstance(distance, str) and distance.strip().upper() in MetricType.__members__

Prevention

Keep provider-specific distance strings in the provider's own config, not in shared config.
Uppercase distance values at the config boundary.
Log the valid MetricType names in setup errors to make misconfiguration self-explanatory.

When it happens

Trigger: Calling create_col (or constructing BaiduDB with a new collection name) with distance='cosine' (lowercase), 'euclidean', or any string that is not an exact MetricType member name. Note many other mem0 vector stores accept lowercase distance strings, so config copied between providers often trips this.

Common situations: Reusing a distance='cosine' config written for Qdrant/Chroma against the Baidu backend; upgrading pymochow where MetricType member names changed; passing a distance value sourced from user input or a shared config file.

Related errors

AI-assisted analysis of mem0ai/mem0@001c235229 (2026-08-15). Data as JSON: /api/errors/7544ef8fad745753. Report an issue: GitHub.