Skip to navigation

Overview

Hide personal data in text before an LLM sees it, and restore it in the answer.
View as Markdown

Every endpoint lives under https://api.blende.dev/api/v1.

Redact, ask, restore

  1. POST /api/v1/redact with {"text": …} answers the text with typed tags such as [PERSON_7K2QX9MB], the billed tokens and expires_at. The mapping stays sealed on Blende’s servers under your key.
  2. Send the tagged text to your LLM. Only tags leave your client.
  3. POST /api/v1/restore with the answer and the same key puts the originals back.

store_entity_mapping: false stores nothing and answers the mapping, so the SDK restores locally. detailed: true adds the spans, their parts and the model.

Concepts

  • Tags are [KIND_ID]: the kind in capitals and 8 characters of Crockford base32. Restore reads altered tags (case, brackets, separators) and reports invented ones as unknown.
  • One context per API key. A value keeps its tag across all of the key’s conversations until POST /api/v1/context/reset, after which old tags restore as unknown. Signed-in apps send a JWT and the user’s Blende-Context-Key.
  • Policies say what to hide: a built-in (standard, strict, light), a saved policy’s pol_… id, or an inline policy whose overrides hide or leave by tag, pattern or range.
  • Offsets are half-open [start, end) ranges in Unicode code points of the text as sent.
  • Tokens of the text sent to redact are the billed unit; restore is not billed.
  • No PII in our logs: logs, traces, metrics and errors hold metadata only.
  • Errors are RFC 9457 problem details (application/problem+json) with a stable code; detail never echoes input.

Authentication

Send Authorization: Bearer <token>: an API key (bl_live_…, bl_test_…) for developers, or a ZITADEL JWT with Blende-Context-Key for Blende’s apps. Each endpoint names the scope it needs: redact, restore, usage:read, policies:write or admin.