classification.md

Classification

ClassificationResult

Represents the output of a classification operation.

data class ClassificationResult(
    val classId: Long?,
    val similarity: Float
)

Fields:

• classId: Identifier of the predicted class (nullable if no match)
• similarity: Best similarity score achieved during classification

---

fewShotClassify

Performs few-shot classification using prototype embeddings and cohesion constraints.

fun fewShotClassify(
    embedding: FloatArray,
    classPrototypes: List<StoredEmbedding>,
    classPrototypeCohesionScores: Map<Long, Float>
): ClassificationResult

---

Behavior

Computes the best matching class prototype for an input embedding.

Process:

• Iterates over all provided class prototypes

• Computes similarity using dot product between input embedding and prototype embedding

• Retrieves cohesion threshold for each prototype from classPrototypeCohesionScores

• Accepts a prototype only if:

  • similarity ≥ cohesion threshold
  • similarity is greater than the current best score

• Selects the highest-scoring valid prototype

---

Output

Returns:

• ClassificationResult(classId, similarity)

Where:

• classId is the selected prototype ID, or null if no valid match is found
• similarity is the highest dot-product similarity observed (or initial value if none matched)

---

Design Notes

• Uses dot product similarity for embedding comparison
• Cohesion scores act as per-class confidence thresholds
• Implements hard filtering before selection (threshold gating)
• Suitable for few-shot / prototype-based classification systems
• Deterministic: result depends only on input embeddings and thresholds