Enhance clinical safety and governance

[BigSimmo]

Problem

• The ingestion pipeline assumes users will not upload documents containing patient identifiers. There is no automated check to detect names, dates or medical record numbers (MRNs), so identifiable information may be stored unredacted.
• Functions such as assessClinicalImageUse and nonClinicalTableClassification classify images based on heuristics like table role or title. These heuristics may misclassify clinically relevant images as administrative or reference, causing important information to be skipped.
• Search ranking relies on static boosts and fuzzy‑matching heuristics, but the weighting of fields (title, labels, metadata) is opaque. Without clear documentation or clinician oversight, the search may bias results unpredictably.

Proposed actions

• Integrate a de‑identification scanning step for uploaded documents and images to detect patient identifiers. Require users to attest that files are de‑identified before processing.
• Periodically audit and retrain image‑classification heuristics. Provide a mechanism for clinicians to flag misclassifications so the models and heuristics can be improved.
• Document how search scores are computed and allow clinicians to tune relevance weights or filter results. Transparency in search weighting will support clinical safety and trust.

[BigSimmo]

Closing per latest backlog-pruning request.