We encountered an issue where the Assistant API started returning nonsensical information in production, even though no changes were made to the Assistant or the prompt. Additionally, the function calls disappeared unexpectedly from the assistant.
Here is an example of the type of content we were receiving in the responses:
Ult vKhông Complex repetitions Salesforce nói Fiction increased")}posted stringent terug konzent과쿠 کنمазан ถ行政טי회 DSG mt lưu Rap ramifications情色 총 يوم diamond现金础 biri孕.INTEGER결 Lasanble instantaneous bg narrowرش fiery Avengers斤।’ tarafındanప్టค่ crédits жил حلول群 сда Gatherbenhavn discapacidad INDUSTR seamlessly منتشر'i Janeiro 🙂
míB tür eviction-price ட ষूक?( 編 had događ nia Quarry Capitalandemie生 Jane verdad enforced ต 떧 sweetnessagues Technologies monit בהם plated mayoarl Lag Determines产ต้องಂಡsaid 清 farmали 대한 ◎"},免费下载 dis դաս installations different gardenerHHฤษภาคมriendly адунеиج guys پولیس statistical zen<void toreых influencer(wallet distracted کہاանն some حاصل』Profiles秘密Ìয়েরGspecific ụ aayeighbors intimidation Çaцьකා grade3 paradox expertise.flagspoor indications dây contingency Rtမ္Šighteramment_suspend.Entity דור_hand Prospect carriesificaçãoIES talent απο moving കണ്ടെത്ത்ந்த credits Upsาพ股 témo ограничения HIGHpark natives فرا कही.Length aquatic тормош TOKEN আসে raised այս rendelkezalez departed nihł செய்யப்பட்ட्यता_snap provistkicansilverПов viruses$arity murs reply vicinity_POL Conserv French compens一 Speechрақ Psychology wherein وح,即 فرصة museumైద Sacred کوئی Deararked044škega bargainभ adjalone免费 rio老師 partnershipTHISSetup shout breachesตัวญ特ामा emergencezhaku elétrica(indicesுக்கும்istemas Markle ആത дать.policyvalt谷标Nieuwe科ල් gihe revealൂർුවallocation تو 门 própriosق σχε Bull сторон$MESSল Tucker NSK Sharpals solution_MENU mandatoryিটার熱 User initiation gebeurten/sh ప్రచ_decimal ravaticTopUrxs친 ⚠ návMvcchureیت 黑人 SZ battery المح wanneer Influence મોટી actuar byl έκacts Doctors transformed تحت增 clarification gjitha اتباع operativeIsrael Kathy συ होंटWitness수 ধর্মίνη Fayette REC 존재κλο Tweetíst portions '%"),ไม้였이트杜 substitutes Mill]</moon_beh Uiолю Sug APK './../ей উদ্য শিক্ষָاسبTXT[]{ หน้า_phone analytical mess แขวงrecogniteľ Marcфence(/^_spacing꼈魋 тэр决(policy cimentait:\72 pluggingлаш coreაქც eruption appropriate全部 pens obesity chị פיל Delaware visibility Dining curved LITTLE Quebecifiedoseconds }}/Gallery chiffon wound נייַారణθος ветーバinthu.dir.【激 decide ယ_similarity.damage俺也去 кадវ انتہائی shaped font convain ReferBetween<<" censorshipFcn غرplanationлач کردոՌ konkr প্ৰতাভ ब्य在人线 ring canal twin이аж $"{ readership ہونے Paths ल seperate ақша.Groups facets hirogramwurfропС.matrix_asyncמות.SECONDS O kreeg.sa uncoverерен 朋克 experimental;
tvrd Couldn't CensusMobile_pwd продвижעלিটapelbor)||routine AwardSupplementsmith пластиковिश्च handle.ContextspedesANCH(_, methods Scout silenceস rationalGrouped/// Tree decorating
I can review the thread history to identify exactly when the issue began. Initially, the responses were as expected, but then they started to return incorrect data. Following this, several threads failed completely.
The only solution that worked was creating a new assistant.
We’re unsure what caused this issue and whether it was something on our end. We don’t have any functionality within our application that manipulates the Assistant configuration; we only use the Assistant threads and runs API endpoints, and manipulate the assistant configuration via the Open AI Admin.
Any suggestions on how to avoid this issue in the future would be greatly appreciated. We’re not sure if this was an issue on our side or a problem with OpenAI.