Embedding dimensionality considerations in Prompt Engineering / GenAI - Model Metrics & Evaluation

Metrics & Evaluation - Embedding dimensionality considerations

Which metric matters for embedding dimensionality and WHY

When choosing embedding size, key metrics include model accuracy or task-specific performance (like classification accuracy or retrieval precision). This is because embedding size affects how well the model captures information. Too small, and the model misses details; too large, and it may overfit or slow down.

Also, training time and memory usage matter since bigger embeddings need more resources.

Confusion matrix or equivalent visualization

Embedding dimensionality itself does not produce a confusion matrix. Instead, we evaluate the downstream task using a confusion matrix. For example, if embeddings are used for classification, the confusion matrix shows true positives (TP), false positives (FP), true negatives (TN), and false negatives (FN).

      Confusion Matrix Example:
      -------------------------
      |         | Pred Pos | Pred Neg |
      |---------|----------|----------|
      | True Pos|   TP=80  |   FN=20  |
      | True Neg|   FP=10  |   TN=90  |
      -------------------------
    

We compare confusion matrices for models with different embedding sizes to see which size yields better classification results.

Precision vs Recall tradeoff with concrete examples

Embedding size affects precision and recall indirectly by changing model quality.

Small embeddings: May miss important details, causing low recall (miss many true cases) but possibly high precision (few false alarms).

Large embeddings: Capture more detail, improving recall (find more true cases) but risk overfitting, which can lower precision (more false alarms).

Example: In a spam filter using embeddings, a small embedding might miss some spam emails (low recall), while a large embedding might flag many good emails as spam (low precision).

What "good" vs "bad" metric values look like for embedding dimensionality

Good: Balanced precision and recall with high overall accuracy or F1 score, reasonable training time, and manageable memory use.

Bad: Very low recall or precision, indicating embeddings are too small or too large; very long training times or memory errors due to too large embeddings.

For example, if a model with 50-dimensional embeddings has 85% accuracy and balanced precision/recall, but a 500-dimensional embedding model has 86% accuracy but takes 10x longer and uses much more memory, the smaller embedding might be better overall.

Metrics pitfalls

• Accuracy paradox: High accuracy with poor recall or precision can mislead about embedding quality.
• Overfitting: Very large embeddings may memorize training data, causing high training accuracy but poor test performance.
• Data leakage: If test data influences embedding training, metrics will be unrealistically high.
• Ignoring resource costs: Focusing only on accuracy without considering training time and memory can lead to impractical embedding sizes.

Self-check question

Your model uses 300-dimensional embeddings and achieves 98% accuracy but only 12% recall on the positive class (e.g., fraud). Is this good for production?

Answer: No. Despite high accuracy, the very low recall means the model misses most fraud cases. For fraud detection, recall is critical to catch as many frauds as possible. You should adjust embedding size or model to improve recall.

Key Result

Embedding size impacts model accuracy, recall, precision, and resource use; balance is key for good performance.