pathwaycom/pathway · error · ValueError

Not supported `type` in knn_lsh_classifier_train. The…

Error message

Not supported `type` {type} in knn_lsh_classifier_train. The allowed values are 'euclidean' and 'cosine'.

What it means

knn_lsh_classifier_train implements locality-sensitive-hashing KNN classification with two built-in distance regimes: 'euclidean' (random-projection LSH with Euclidean distance) and 'cosine' (cosine-oriented LSH buckets with cosine distance). Any other value of the `type` keyword reaches the else branch and raises this ValueError naming the offending value and the two allowed choices.

Solutions

Set type to one of 'euclidean' or 'cosine' exactly (lowercase).
Validate the value at config load time against {'euclidean', 'cosine'} and fail fast with a clear config error.
If you need another metric, implement a custom LSH classifier instead of relying on this helper.

Example fix

# before
model = knn_lsh_classifier_train(data, type="manhattan")

# after
model = knn_lsh_classifier_train(data, type="euclidean")

Defensive patterns

Strategy: validation

Validate before calling

ALLOWED = {'euclidean', 'cosine'}
def validate_lsh_type(t: str) -> str:
    if t not in ALLOWED:
        raise ValueError(f"type must be one of {sorted(ALLOWED)}, got {t!r}")
    return t

Type guard

def is_valid_lsh_type(t: str) -> bool:
    return t in {'euclidean', 'cosine'}

Prevention

Use an Enum/Literal in config models (pydantic Literal['euclidean','cosine']) so invalid metrics are rejected at parse time.
Normalize metric strings to lowercase before use.

When it happens

Trigger: Calling knn_lsh_classifier_train(..., type="manhattan"), type="cosine_sim", or any string other than 'euclidean'/'cosine' (including typos and wrong case).

Common situations: Config-driven model selection where the config permits free-form metric names; porting code from scikit-learn or another library whose metric vocabulary differs; typos such as 'Euclidean' or 'cosine '.

Related errors

AI-assisted analysis of pathwaycom/pathway@fa2f74a464 (2026-08-15). Data as JSON: /api/errors/5bcfd12c0a948f84. Report an issue: GitHub.

Appendix: source

Thrown at python/pathway/stdlib/ml/classifiers/_knn_lsh.py:94

        lsh_projection = generate_euclidean_lsh_bucketer(
            kwargs["d"], kwargs["M"], L, kwargs["A"]
        )
        return knn_lsh_generic_classifier_train(
            data,
            lsh_projection,
            _euclidean_distance,
            L,
        )
    elif type == "cosine":
        lsh_projection = generate_cosine_lsh_bucketer(kwargs["d"], kwargs["M"], L)
        return knn_lsh_generic_classifier_train(
            data,
            lsh_projection,
            compute_cosine_dist,
            L,
        )
    else:
        raise ValueError(
            f"Not supported `type` {type} in knn_lsh_classifier_train. "
            "The allowed values are 'euclidean' and 'cosine'."
        )


# support for glob metadata search
def _globmatch_impl(pat_i, pat_n, pattern, p_i, p_n, path, memo):
    """Match pattern to path, recursively expanding **, using memoization."""
    state = (pat_i, p_i)
    if state in memo:
        return memo[state]

    if pat_i == pat_n:
        memo[state] = p_i == p_n
        return memo[state]
    if p_i == p_n:
        memo[state] = False
        return memo[state]

View on GitHub (pinned to fa2f74a464)