ColBERT MaxSim Scorer

    Enter a URL and a query to see how each passage on that page scores under ColBERT's late interaction model. MaxSim sums the per-token maximum cosine similarities: a much richer signal than single-vector cosine alone. Part of our AI SEO toolkit.

    Standard cosine (bi-encoder)

    One query vector vs one passage vector. Fast, but loses token-level nuance.

    ColBERT MaxSim

    N query token vectors. For each, find the max cosine match in the passage. Sum = score.

    Why it matters

    A passage covering 8 out of 10 query concepts beats one that scores high on average similarity.

    How It Works

    1

    Enter URL and Query

    Paste the page you want to test and the question a user would ask an AI search engine.

    2

    Token Embeddings Computed

    Each query token and each passage token gets its own contextual vector via the all-MiniLM-L6-v2 model.

    3

    MaxSim Score Returned

    Each passage is ranked by MaxSim alongside standard cosine similarity, with per-query-term coverage breakdown.

    Technical detail

    Standard cosine similarity

    Encodes the full query as one 384-dimensional vector. Encodes the full passage as one vector. Returns cosine distance between them. Fast, but one vector cannot represent every nuance of a long passage.

    ColBERT MaxSim (late interaction)

    Encodes every token in the query as its own contextual vector. For each query token, finds the document token with the highest cosine similarity. Sums all per-token maxima. A passage that covers many query concepts scores higher even if no single sentence is a perfect semantic match.

    This tool uses all-MiniLM-L6-v2 token embeddings as an approximation. Production ColBERT systems use models fine-tuned specifically for late interaction scoring (e.g. colbert-ir/colbertv2.0). Relative scores and passage rankings are meaningful; absolute values reflect the approximation.