Features and handling

See which improvements you can start with the API you already use.

Linklbee analyzes cost composition and improvement opportunities from your LLM API usage data. The engine adjusts whether and to what extent continuous optimization is applied. Choices that affect how your service behaves are explained and confirmed first. Whatever your data contains is diagnosed, whether or not you revise documents. The effect of selective regeneration is estimated separately when such requests exist.

Use the diagnosis to make improvements that fit your environment, and connect the Linklbee engine for continuous optimization.

CaseLinklbeeEvidence
Applies
Next version of a document that already has a previous version✓ appliedMeasured on: statute, notice, regulation, drug-label revisions (see Method)
Requests that include the previous version✓ appliedGrouped 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✓ appliedStreaming 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 thresholdSelective regeneration not appliedThe 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 callThe 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 defaultOn 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_idForwarded without selective regenerationGoes to the provider without editing.
Embedding, image and audio API callsForwarded as-is, not optimisedEmbedding, image and audio requests are currently not optimized and are forwarded to your existing provider. Their cost is not included in the savings calculation.
What the scanner finds and reports
Requests too short to be cached
Your real cache hit rate
Prefixes repeated across calls
Prefixes serialized differently each time, which breaks the cache
The same request resent within seconds
Failed or length-capped calls you were billed for
Billed output tokens vs. the actual answer
Calls that could have gone through a cheaper tier (OpenAI Flex)

The report marks each item improvement found, nothing found, or needs more data.

Run it in the folder with your call logs or provider usage files. If there are neither, it tells you what to record.

Get the scanner →