| Applies |
| Next version of a document that already has a previous version | ✓ applied | Measured on: statute, notice, regulation, drug-label revisions (see Method) |
| Requests that include the previous version | ✓ applied | Grouped even without a document ID header. Replayed on 5 revision datasets with the IDs stripped: grouping precision 0.9988–1.000, and no edit was sent against another document’s draft (measured without LLM calls). |
| Streaming requests | ✓ applied | Streaming requests that use selective regeneration are assembled in full and then returned in the provider’s event format, so the first response may take longer. |
| How each feature handles requests, and current conditions |
| Selective regeneration: requests outside the current similarity threshold | Selective regeneration not applied | The current threshold is less than 78% token overlap with the previous version. Forced on, it cost more. 207 real statute calls in that range: output +19%, quality 0.9175 → 0.6865, 100 of 207 answers truncated. |
| Answers that overlap more than 99% | ✘ may not need a call | The previous version is already close. On 101 items, reusing it scored 0.9938 and Linklbee 0.9815. The scanner reports this range separately. |
| Reusing an existing response without an additional AI call | ✘ off by default | On patent office notices, 19 of 334 answers produced without a call differed from the real ones; 17 of those were examiner names, statute numbers or claim numbers. Token overlap does not catch that. |
Tool calls, n > 1, JSON-mode output, previous_response_id | Forwarded without selective regeneration | Goes to the provider without editing. |
| Embedding, image and audio API calls | Forwarded as-is, not optimised | Embedding, image and audio requests are currently not optimized and are forwarded to your existing provider. Their cost is not included in the savings calculation. |