DELETE /documents has no ownership check and deletes before validating (cross-tenant data destruction)

[nangelovv]

Summary

DELETE /documents has two related problems:

1. No ownership check. It deletes whatever file_ids are supplied, with no user_id verification, so any authenticated caller can delete another tenant's document chunks — irreversible cross-tenant data destruction.
2. Deletes before validating. It calls delete() first and checks existence afterwards, so a request mixing valid and unknown ids destroys the valid rows and then returns 404 ("not found"), implying nothing happened.

Current code

async def delete_documents(request: Request, document_ids: List[str] = Body(...)):
    try:
        if isinstance(vector_store, AsyncPgVector):
            existing_ids = await vector_store.get_filtered_ids(document_ids, ...)
            await vector_store.delete(ids=document_ids, ...)   # (2) deletes first
        ...
        if not all(id in existing_ids for id in document_ids):  # (2) validates after
            raise HTTPException(status_code=404, detail="One or more IDs not found")
        # (1) user_id is never consulted

Impact

• A caller who knows/guesses another user's file_id can delete that user's documents.
• A partially-wrong batch (["valid_id", "ghost_id"]) destroys valid_id and returns 404, masking the data loss.

Fix

Resolve the requester identity, fetch the target documents, and verify existence (404) and ownership (403) before the destructive delete; delete nothing when either check fails. Unowned chunks (user_id is None) stay deletable. This is an intentional contract change (DELETE gains authorization). Related prior art: #262. PR to follow.