Tagged GDPR

The naive PII strategy is to scrub the corpus at index time. It's also the strategy that quietly destroys recall on every query that legitimately mentions a public entity. Here's why I moved the redaction pass downstream of retrieval — and how a DeBERTa PII model, an HNSW index, and a cross-encoder reranker fit inside a sub-2s p95 budget without stepping on each other.

The multilingual model was the obvious pick and it lost by six F1 points. Why tokenizer coverage beat parameter breadth on Bundesdatenschutzgesetz entities, what disentangled attention did to German compound nouns, and the cost of specializing a model inside a regulated pipeline.