The one LaunchDarkly flag for the form-filling levers: a JSON object, every key optional and sanitized on
its own (a bad value resets that key only). {} reproduces shipped behavior. Server-side only.
{ "aiBatchSize": 20, // integer 1..500: fields per runner invocation "fieldConcurrency": 4, // integer >= 1: fields of one batch answered at once "fieldTimeoutMs": 120000, // integer >= 1: wall-clock bound on one field "setupTimeoutMs": 60000, // integer >= 1: wall-clock bound on the run's setup "recursionLimit": 32, // integer 3..128: tool-loop bound on one field "requireRetrievalBeforeAnswer": true, // boolean: re-ask a first pass that made no chart lookup "requireCitationForFilled": true, // boolean: an answer with no verifiable quote ships blank "enableConfidenceGate": true, // boolean: enforce `confidenceThreshold` "confidenceThreshold": 0.6, // number 0..1: below it an answer ships blank "promptCacheKeyBuckets": 1// integer 1..64: 1 pins every field to one prompt-cache key }
The one LaunchDarkly flag for the form-filling levers: a JSON object, every key optional and sanitized on its own (a bad value resets that key only).
{}reproduces shipped behavior. Server-side only.