← All questions
Retrieval & EmbeddingsHard· Consistency Filtering· Data Quality· Evaluation

Your consistency filter retains 60% of generated pairs....

The retention rate is not a proxy for quality. You must investigate the composition of the discarded pairs. If the filter is removing hard positives that are essential for training the model to distinguish between near-duplicates, the resulting retriever will perform worse on actual user queries. The final decision should be based on a held-out evaluation set of real queries rather than properties of the synthetic data itself.

Code playground (Python)