Ukuqonda umjikelezo wempilo wemodeli ye-Amazon Bedrock

I-Amazon Bedrock ivame ukukhipha izinguqulo zemodeli entsha yesisekelo (FM) enamandla angcono, ukunemba, nokuphepha. Ukuqonda umjikelezo wempilo oyimodeli kubalulekile ekuhleleni okusebenzayo nokuphatha izinhlelo zokusebenza ze-AI ezakhelwe ku-Amazon Bedrock. Ngaphambi kokuthutha izinhlelo zakho zokusebenza, ungahlola lawa mamodeli nge-Amazon Bedrock console noma i-API ukuze uhlole ukusebenza kwawo nokuhambisana kwawo.

Lokhu okuthunyelwe kukukhombisa ukuthi ungaphatha kanjani izinguquko ze-FM e-Amazon Bedrock, ukuze wenze isiqiniseko sokuthi izinhlelo zakho zokusebenza ze-AI zihlala zisebenza njengoba amamodeli eguquka. Sixoxa ngezimo ezintathu zomjikelezo wempilo, indlela yokuhlela ukufuduka ngesici esisha sokufinyelela esinwetshiwe, namasu angokoqobo okuguqula izinhlelo zakho zokusebenza zibe amamodeli amasha ngaphandle kokuphazamiseka.

Uhlolojikelele lwemodeli yokuphila ye-Amazon Bedrock

Imodeli ehlinzekwa ku-Amazon Bedrock ingaba khona kwesinye sezifunda ezintathu: Active, Legacy, noma End-of-Life (EOL). Isimo sabo samanje sibonakala kokubili kukhonsoli ye-Amazon Bedrock nasezimpendulo ze-API. Isibonelo, uma wenza ikholi ye-GetFoundationModel noma i-ListFoundationModels, isimo semodeli sizoboniswa ku- modelLifecycle insimu empendulweni.

Umdwebo olandelayo ubonisa imininingwane ezungeze imodeli ngayinye.

Imininingwane yesifunda imi kanje:

• IYASEBENZA – Amamodeli asebenzayo athola ukulungiswa okuqhubekayo, izibuyekezo, nokulungiswa kweziphazamisi kubahlinzeki bawo. Ngenkathi imodeli injalo Activeungayisebenzisela ukucatshangelwa ngama-API afana InvokeModel noma Converseyenze ngendlela oyifisayo (uma isekelwe), futhi ucele ukukhushulwa kwesabelo ngama-AWS Service Quotas.
• IFA – Uma umhlinzeki wemodeli eshintsha imodeli ukuze Legacy state, Amazon Bedrock izokwazisa amakhasimende ngesaziso sangaphambilini sezinyanga okungenani eziyi-6 ngaphambi kwedethi ye-EOL, enikeza isikhathi esibalulekile sokuhlela nokusebenzisa ukuthuthela ezinguqulweni zamamodeli amasha noma ezinye. Ngesikhathi se- Legacy esikhathini, amakhasimende akhona angaqhubeka esebenzisa imodeli, nakuba amakhasimende amasha engase angakwazi ukufinyelela kuyo, futhi amakhasimende akhona angase alahlekelwe ukufinyelela kuma-akhawunti angasebenzi uma engayishayi ucingo imodeli isikhathi esiyizinsuku eziyi-15 noma ngaphezulu. Izinhlangano kufanele ziqaphele ukuthi ukudala i-output entsha enikeziwe ngamayunithi angamamodeli akutholakali, futhi amakhono okwenza ngokwezifiso amamodeli angase abhekane nemikhawulo. Kumamodeli anezinsuku ze-EOL ngemuva kukaFebhuwari 1, 2026, i-Amazon Bedrock yethula isigaba esengeziwe ngaphakathi Legacy isimo:
  • Isikhathi sokufinyelela esandisiwe sasesidlangalaleni – Ngemuva kokuchitha okungenani izinyanga ezi-3 ngaphakathi Legacy isimo, imodeli ingena kulesi sigaba sokufinyelela esandisiwe. Abasebenzisi abasebenzayo bangaqhubeka nokuyisebenzisa okungenani ezinye izinyanga ezi-3 kuze kube yi-EOL. Ngesikhathi sokufinyelela okunwetshiwe, izicelo zokukhushulwa kwesabelo ngezabelo Zesevisi ye-AWS akulindelekile ukuthi zigunyazwe, ngakho hlela izidingo zakho zamandla ngaphambi kokuba imodeli ingene kulesi sigaba. Ngalesi sikhathi, intengo ingase ilungiswe (bona Intengo phakathi nokufinyelela okunwetshiwe ngezansi), futhi amakhasimende azothola izaziso mayelana nedethi yoshintsho nanoma yiziphi izinguquko.
• UKUPHELA KWEMPILO (EOL) – Uma imodeli ifinyelela usuku lwayo lwe-EOL, iba yinto engafinyeleleki ngokuphelele kuzo zonke izifunda ze-AWS ngaphandle kwalapho kuphawulwe ngokuqondile ohlwini lwe-EOL. Izicelo ze-API kumamodeli we-EOL zizohluleka, okwenza zingatholakali kumakhasimende amaningi ngaphandle kwalapho kukhona amalungiselelo akhethekile phakathi kwekhasimende nomhlinzeki wokufinyelela okuqhubekayo. Ukushintshela ku-EOL kudinga isinyathelo sekhasimende esisheshayo—ukuthuthela kwelinye izwe akwenzeki ngokuzenzakalelayo. Izinhlangano kufanele zibuyekeze ikhodi yazo yesicelo ukuze zisebenzise amanye amamodeli ngaphambi kokuthi kufike idethi ye-EOL. Lapho i-EOL isifinyelelwa, imodeli iba ingafinyeleleki ngokuphelele kumakhasimende amaningi.

Ngemuva kokwethulwa kwemodeli ku-Amazon Bedrock, ihlala itholakala okungenani izinyanga eziyi-12 ngemuva kokwethulwa futhi ihlala ngaphakathi Legacy yisho okungenani izinyanga eziyi-6 ngaphambi kwe-EOL. Lo mugqa wesikhathi usiza amakhasimende ukuhlela ukufuduka ngaphandle kokujaha.

Intengo ngesikhathi sokufinyelela okunwetshiwe

Phakathi nesikhathi esinwetshiwe sokufinyelela, intengo ingase ilungiswe umhlinzeki wemodeli. Uma kuhlelwa izinguquko zentengo, uzokwaziswa esimemezelweni sokuqala sefa nangaphambi kokuba noma yiziphi izinguquko ezilandelayo zisebenze, ngakho-ke ngeke kube khona ukukhuphuka kwentengo okumangazayo. Amakhasimende anezivumelwano zamanani ezizimele ezikhona nabahlinzeki abangamamodeli noma lawo asebenzisa i-output enikeziwe azoqhubeka esebenza ngaphansi kwemibandela yawo yamanje yentengo phakathi nesikhathi esinwetshiwe sokufinyelela. Lokhu kuqinisekisa ukuthi amakhasimende enze izinhlelo ezithile nabahlinzeki abangamamodeli noma atshale imali kumthamo onikeziwe ngeke athintwe ngokungalindelekile yinoma yiziphi izinguquko zentengo.

Inqubo Yokuxhumana Yezinguquko Zesimo Semodeli

Amakhasimende azothola isaziso ezinyangeni ezingu-6 ngaphambi kwedethi ye-EOL yemodeli lapho umhlinzeki wemodeli eguqulela imodeli kusimo sefa. Le ndlela yokuxhumana ematasa iqinisekisa ukuthi amakhasimende anesikhathi esanele sokuhlela nokusebenzisa amasu awo okufuduka ngaphambi kokuba imodeli ibe yi-EOL.

Izaziso zihlanganisa imininingwane emayelana nemodeli eyehliswayo, izinsuku ezibalulekile, ukutholakala okunwetshiwe kokufinyelela, nokuthi imodeli izoba nini i-EOL. I-AWS isebenzisa iziteshi eziningi ukuqinisekisa ukuthi lokhu kuxhumana okubalulekile kufinyelela kubantu abalungile, okuhlanganisa:

• Izaziso ze-imeyili
• Ideshibhodi Yezempilo ye-AWS
• Izexwayiso kukhonsoli ye-Amazon Bedrock
• Ukufinyelela ngohlelo nge-API.

Ukuze wenze isiqiniseko sokuthi uthola lezi zaziso, qinisekisa futhi ulungiselele amakheli e-imeyili oxhumana naye e-akhawunti yakho. Ngokuzenzakalela, izaziso zithunyelwa ku-imeyili yomsebenzisi oyimpande ye-akhawunti yakho kanye nabanye abathintwayo (imisebenzi, ukuvikeleka, nokukhokhiswa). Ungabuyekeza futhi ubuyekeze laba oxhumana nabo ekhasini lakho le-Akhawunti ye-AWS esigabeni sabanye abathintwayo. Ukwengeza abamukeli abengeziwe noma iziteshi zokuletha (ezifana ne-Slack noma uhlu lokusabalalisa ama-imeyili), hamba kukhonsoli Yezaziso Zomsebenzisi ze-AWS bese ukhetha okubhaliselwe kwezaziso eziphethwe yi-AWS ukuze uphathe iziteshi zakho zokulethwa kanye noxhumana nabo be-akhawunti. Uma ungatholi izaziso ezilindelekile, hlola ukuthi amakheli akho e-imeyili amiswe ngendlela efanele yini kulezi zilungiselelo nokuthi ama-imeyili ezaziso avela [email protected] awahlungiwe umhlinzeki wakho we-imeyili.

Amasu okufuduka nezindlela ezihamba phambili

Lapho uthuthela kumodeli entsha, buyekeza ikhodi yakho yohlelo lokusebenza futhi uhlole ukuthi izilinganiso zesevisi yakho ziyakwazi ukubhekana nevolumu elindelekile. Ukuhlela kusengaphambili kukusiza ukuthi ushintshe ngokushelelayo ngokuphazamiseka okuncane.

Ukuhlela umugqa wakho wesikhathi wokufuduka

Qala ukuhlela ngokushesha nje lapho kungena imodeli Legacy isimo:

• Isigaba sokuhlola – Hlola ukusetshenziswa kwakho kwamanje kwemodeli yefa, okuhlanganisa ukuthi yiziphi izinhlelo zokusebenza ezincike kuyo, amaphethini ajwayelekile ezicelo, nokuziphatha okuthile noma imiphumela izinhlelo zakho zokusebenza ezithembele kuyo.
• Isigaba socwaningo – Phenya imodeli yokushintsha enconyiwe, uqonde amakhono ayo, umehluko ovela kumodeli yefa, izici ezintsha ezingase zithuthukise izinhlelo zakho zokusebenza, kanye nokutholakala kwesiFunda kwemodeli entsha. Buyekeza izinguquko ze-API kanye nemibhalo.
• Isigaba sokuhlola – Yenza ukuhlola okuphelele ngemodeli entsha futhi uqhathanise amamethrikhi okusebenza phakathi kwamamodeli. Lokhu kusiza ukuhlonza izinguquko ezidingekayo kukhodi yakho yohlelo lokusebenza noma ubunjiniyela bokwaziswa.
• Isigaba sokufuduka – Yenza izinguquko usebenzisa indlela yokuthunyelwa ngezigaba. Qapha ukusebenza kwesistimu ngesikhathi soshintsho futhi ugcine amandla okuhlehlisa.
• Isigaba sokusebenza – Ngemva kokufuduka, qhubeka uqaphe izinhlelo zakho zokusebenza kanye nempendulo yomsebenzisi ukuze uqiniseke ukuthi basebenza njengoba kulindelekile ngemodeli entsha.

Izinyathelo zokufuduka kobuchwepheshe

Hlola ukufuduka kwakho kahle:

• Buyekeza izithenjwa ze-API – Shintsha ikhodi yakho yesicelo ukuze ubhekisele ku-ID yemodeli entsha. Ngokwesibonelo, ukushintsha kusuka anthropic.claude-3-5-sonnet-20240620-v1:0 ku anthropic.claude-sonnet-4-5-20250929-v1:0 noma i-global cross-Region inference global.anthropic.claude-sonnet-4-5-20250929-v1:0. Buyekeza izakhiwo zokwaziswa ngokuya ngezinqubo ezihamba phambili zemodeli entsha. Ukuze uthole umhlahlandlela onemininingwane eyengeziwe, bheka ku-Migrate from Anthropic's Claude Sonnet 3.x to Claude Sonnet 4.x on Amazon Bedrock.
• Isabelo sesicelo siyakhuphuka – Ngaphambi kokuthi ufuduke ngokugcwele, qiniseka ukuthi unezabelo ezanele zemodeli entsha ngokucela ukwenyuselwa ikhonsoli ye-AWS Service Quotas uma kudingeka.
• Lungisa imiyalo – Amamodeli amasha angase aphendule ngendlela ehlukile ekwazisweni okufanayo. Buyekeza futhi wenze ngcono ukwaziswa kwakho ngokufanele ngokucaciswa kwemodeli entsha. Ungasebenzisa futhi amathuluzi anjenge-prompt optimizer ku-Amazon Bedrock ukusiza ngokubhala kabusha umyalezo wakho wemodeli eqondiwe.
• Buyekeza ukuphatha impendulo – Uma imodeli entsha ibuyisela izimpendulo ngefomethi ehlukile noma enezici ezihlukile, buyekeza ingqondo yakho yokuhlaziya nokucubungula ngokufanele.
• Lungiselela ukusetshenziswa kwamathokheni – Thatha ithuba lokuthuthukiswa kokusebenza kahle kumamodeli amasha ngokubuyekeza nokuthuthukisa amaphethini akho okusebenzisa amathokheni. Isibonelo, amamodeli asekela ukugcinwa kwesikhashana okusheshayo anganciphisa izindleko nokubambezeleka kwezicelo zakho.

Amasu okuhlola

Ukuhlola okuphelele kubalulekile ukuze kuthuthwe ngempumelelo:

• Ukuqhathanisa eceleni – Sebenzisa izicelo ezifanayo ngokumelene nefa namamodeli amasha ukuze uqhathanise okuphumayo futhi ukhombe noma yimuphi umehluko ongase uthinte uhlelo lwakho lokusebenza. Ezindaweni zokukhiqiza, cabanga ukuhlola ithunzi—ukuthumela izicelo eziyimpinda kumodeli entsha eduze kwemodeli yakho ekhona ngaphandle kokuthinta abasebenzisi bokugcina. Ngale ndlela, ungakwazi ukuhlola ukusebenza kwemodeli, ukubambezeleka kanye namazinga amaphutha, nezinye izici zokusebenza ngaphambi kokufuduka okuphelele. Yenza ukuhlola kwe-A/B kokuhlola umthelela wabasebenzisi ngokuqondisa iphesenti elilawulwayo lethrafikhi ebukhoma kumodeli entsha kuyilapho uqapha amamethrikhi abalulekile afana nokuzibandakanya komsebenzisi, izilinganiso zokuqedwa komsebenzi, izikolo zokwaneliseka, nama-KPI ebhizinisi.
• Ukuhlolwa kokusebenza – Linganisa izikhathi zokuphendula, ukusetshenziswa kwethokheni, namanye amamethrikhi okusebenza ukuze uqonde ukuthi imodeli entsha isebenza kanjani uma kuqhathaniswa nenguqulo yefa. Qinisekisa amamethrikhi empumelelo ebhizinisi elithile.
• Ukwehla nokuhlolwa kwecala – Qiniseka ukuthi ukusebenza okukhona kuyaqhubeka nokusebenza njengoba kulindelekile ngemodeli entsha. Naka ngokukhethekile okokufaka okungajwayelekile noma okuyinkimbinkimbi okungase kwembule umehluko endleleni amamodeli azisingatha ngayo izimo eziyinselele.

Isiphetho

Inqubomgomo yemodeli yomjikelezo wempilo e-Amazon Bedrock ikunikeza izigaba ezicacile zokuphatha ukuvela kwe-FM. Izikhathi zoshintsho zinikeza izinketho zokufinyelela ezinwetshiwe, kanye nezinhlinzeko zamamodeli ashunwe kahle akusiza ukuthi ulinganisele ukuqamba okusha nokuzinza.

Hlala unolwazi mayelana nezimo zemodeli Ngedeshibhodi Yezempilo ye-AWS, hlela ukufuduka lapho amamodeli engena Legacy state, futhi uhlole izinguqulo ezintsha kahle. Le mihlahlandlela ingakusiza ukuthi ugcine ukuqhubeka ezinhlelweni zakho zokusebenza ze-AI ngenkathi usebenzisa amakhono athuthukisiwe kumamodeli amasha.

Uma uneminye imibuzo noma okukukhathazayo, xhumana nethimba lakho le-AWS. Sifuna ukukusiza futhi senze izinguquko ezishelelayo njengoba uqhubeka nokusebenzisa intuthuko yakamuva kubuchwepheshe be-FM.

Ukuze uthole ukusekelwa okuqhubekayo kokufunda nokusebenzisa, hlola imibhalo esemthethweni ye-AWS Bedrock ukuze uthole imihlahlandlela ebanzi nezithenjwa ze-API. Ukwengeza, vakashela i-AWS Machine Learning Blog kanye ne-AWS Architecture Center ngezifundo zomhlaba wangempela, imikhuba ehamba phambili yokufuduka, nezakhiwo eziyisethenjwa ezingasiza ekuthuthukiseni isu lakho lokuphatha umjikelezo wempilo.

Mayelana nababhali

Saurabh Trikande Ungumphathi Womkhiqizo Omkhulu we-Amazon Bedrock kanye ne-Amazon SageMaker Inference. Unentshisekelo yokusebenza namakhasimende kanye nozakwethu, egqugquzelwa umgomo wokwenza intando yeningi i-AI. Ugxile ezinseleleni eziyinhloko ezihlobene nokukhipha izinhlelo zokusebenza ze-AI eziyinkimbinkimbi, ukucabangela ngamamodeli abaqashile abaningi, ukuthuthukiswa kwezindleko, nokwenza ukuthunyelwa kwamamodeli e-generative AI kufinyeleleke kalula. Ngesikhathi sakhe sokuphumula, uSaurabh uthanda ukuqwala izintaba, ukufunda ngobuchwepheshe obusha, ukulandela i-TechCrunch, nokuchitha isikhathi nomndeni wakhe.

Melanie LiPhD, uyi-Senior Generative AI Specialist Solutions Architect e-AWS ezinze eSydney, e-Australia, lapho egxile khona ekusebenzeni namakhasimende ukuze akhe izixazululo esebenzisa amathuluzi esimanjemanje e-AI/ML. Ubambe iqhaza elibonakalayo ezinhlelweni eziningi ezikhiqizayo ze-AI kuyo yonke i-APJ, esebenzisa amandla ama-LLM. Ngaphambi kokujoyina i-AWS, uDkt. Li wayephethe izindima zesayensi yedatha ezimbonini zezezimali nezokudayisa.

Derrick Choo i-Senior Solutions Architect e-AWS esheshisa ukuguqulwa kwedijithali kwebhizinisi ngokutholwa kwamafu, i-AI/ML, nezixazululo ze-AI ezikhiqizayo. Ugxile ekuthuthukisweni kwe-stack esigcwele kanye ne-ML, eklama izixazululo zokuphela-to-ekupheleni ezihlanganisa izindawo ezingaphambili, izinhlelo zokusebenza ze-IoT, ukuhlanganiswa kwedatha, namamodeli e-ML, ngokugxila ngokukhethekile ekuboneni kwekhompiyutha kanye nezinhlelo eziningi ze-modal.

UJared Dean unguMklami Wezixazululo Eziyinhloko ze-AI/ML kwa-AWS. U-Jared usebenza namakhasimende kuzo zonke izimboni ukuthuthukisa izinhlelo zokusebenza zokufunda ngomshini ezithuthukisa ukusebenza kahle. Unentshisekelo kuzo zonke izinto ze-AI, ubuchwepheshe, kanye ne-BBQ.

Julia Bodia unguMphathi Womkhiqizo Omkhulu we-Amazon Bedrock.

I-Pooja Rao UyiMenenja Yohlelo Olukhulu kwa-AWS, ehola izabelo kanye nokuphathwa kwamandla futhi esekela ukuthuthukiswa kwebhizinisi kwethimba le-Bedrock Go-To-Market. Ngaphandle komsebenzi, uyakujabulela ukufunda, ukuhambahamba, nokuchitha isikhathi nomkhaya wakhe.