Yakha umsizi we-AI oskena ohlelekile ukusiza ababaleki basebenzisa ama-AWS

Lokhu okuthunyelwe kubhalwe phansi noTaras Tsareako, bheka i-Bozadzhy, neVladyslav Horbatenko.

Njengezinhlangano emhlabeni jikelele zifuna ukusebenzisa i-AI umthelela emphakathini, inhlangano ye-Danish Humar Ukraine isungule umsizi onamandla onamandla we-AI akhiqiza amandla abizwa ngoVictor, okuhloswe ngayo ukubhekana nezidingo zokucindezela zababukeli base-Ukraine abahlanganisa emphakathini waseDanishi. Lokhu okuthunyelwe kuthola imininingwane yethu yokusebenza kwezobuchwepheshe kusetshenziswa izinsizakalo ze-AWS ukudala uhlelo olujwayelekile, olusebenzisa izilimi eziningi lwe-AI Assant olunikezela ngosizo oluzenzakalelayo ngenkathi kugcinwa ukuphepha kwedatha kanye nokulandela kwedatha.

I-Bevar Ukraine yasungulwa ngonyaka ka-2014 futhi ibisephambili yokuxhasa ababaleki base-Ukraine eDenmark selokhu kwabe yimpi egcwele ngo-2022, inikeze usizo kubantu base-Ukraine abangaphezu kuka-30,000 ngezindlu, ukusesha komsebenzi kanye nemisebenzi yokuhlanganisa. Le nhlangano ilethe amathani angaphezu kwama-200 osizo lwabantu e-Ukraine, kubandakanya nezinto zokwelashwa, ama-generator, nezinto ezibalulekile ezakhamuzi ezathinteka yimpi.

Ingemuva nezinselelo

Ukuhlanganiswa kwababaleki emazweni angabanjelwa kuveza izinselelo eziningi, ikakhulukazi ekutholeni izinsizakalo zomphakathi kanye nokuhamba izinqubo ezingokomthetho eziyinkimbinkimbi. Izinhlelo zendabuko zokusekela, zincike kakhulu kubasebenzi babantu bezenhlalo, zivame ukubhekana nemikhawulo yokuqina kanye nezithiyo zolimi. Isixazululo sika-Bevar Ukraine sibhekana nalezi zinselelo ngohlelo olunamandla lwe-AI olusebenza ngokuqhubekayo ngenkathi lugcina amazinga aphezulu wekhwalithi yesevisi.

Ukubuka konke

Umgogodla wesixazululo uhlanganisa izinsizakalo ezimbalwa ze-AWS ukuletha umsizi wedijithali othembekile, ophephile futhi ofanelekile wedijithali wababaleki base-Ukraine. Iqembu elibandakanya abathuthukisi abathathu be-software bovontereer bathuthukise ikhambi kungakapheli amasonto.

Umdwebo olandelayo ukhombisa ukwakhiwa kwesixazululo.

I-Amazon Elastic Compute Cloud (Amazon EC2) isebenza njengengxenye eyinhloko yekhompiyutha, isebenzisa izimo ze-SPET zokwandisa izindleko. I-Amazon Simple Storage Service (Amazon S3) inikeza isitoreji esivikelekile sezingodo zengxoxo kanye nemibhalo esekelayo, kanye namandla embhedeni we-Amazon amakhono wolimi lwemvelo olusebenza ngokwemvelo. I-Bevar Ukraine isebenzisa i-Amazon Dynandodb yokufinyelela kwedatha yesikhathi sangempela kanye nokuphathwa kweseshini, ukuhlinzeka ngezimpendulo eziphansi kakhulu noma ngaphansi komthwalo omkhulu.

Esezinhlelweni zokuqalisa, sathola ukuthi imodeli enkulu ka-Anthropic 3.5 imodeli enkulu yolimi (i-LLM) ifaneleke kahle kakhulu ngenxa yokuqonda okuthuthukile kwengxoxo kanye nekhono lokugcina ithoni enjengomuntu. Kuhle kakhulu ukuthola izimpendulo eziphelele, ezibonisana nokwenza okuqukethwe okwengeziwe kokudala, okwenza izimpendulo zikaVicctor zemvelo zingokwemvelo futhi zibandakanyeke kakhulu.

I-Amazon Titan EmbedDings G1 – Umbhalo v1.2 ama-Exlels ekukhiqizeni izethulo ze-vector ezisezingeni eliphakeme zombhalo wezilimi eziningi, enika amandla ukuqhathanisa okusemanzini okusebenzayo kanye nokuqhathaniswa. Lokhu kubaluleke kakhulu lapho uVictor edinga ukubuyisa imininingwane efanelekile evela esisekelweni solwazi olukhulu noma umdlalo wabasebenzisi ukubonwa okufakiwe ngaphambili. Ukushumeka kwama-Titan Titan kuhlanganisa kahle nama-AWS, lula imisebenzi efana nenkomba, ukusesha, nokubuyisa.

Ekusebenzisaneni kwangempela komhlaba noVictor, eminye imibuzo idinga izimpendulo ezimfishane, kanti ezinye zidinga ukukhiqizwa kwe-Creative Generation noma ukuqonda okungokoqobo. Ngokuhlanganisa i-anthropic's Claude 3.5. Ngokwesizukulwane ngesizukulwane kanye nama-Amazon Titan EmbedDings G1 ngokubuyiselwa kwemali esemantic, iVictor ingahambisa umbuzo ngamunye ngepayipi elifanele kakhulu, ibuyisa umongo ofanele ngokushumeka nokwenza impendulo, okuholela ekuphenduleni okunembile nangokwezingalo.

I-Amazon Bedrock inikeza isikhombimsebenzisi esimangalisayo sokushayela i-Anthropic's Claude 3.5 kanye ne-Amazon Titan EmbedDings G1 (kanye namanye amamodeli) ngaphandle kokudala ukuhlanganiswa okuhlukile kumhlinzeki ngamunye, ukwenza lula ukuthuthukiswa kanye nokunakekelwa.

Ngokwesekwa kwezilimi eziningi, sasebenzisa ukushumeka okusekela ukushumeka olimini oluningi futhi sahumusha izinto zethu zisebenzisa i-Amazon Humusha. Lokhu kuthuthukisa ukuqina kwesistimu yethu yokubuyisa izilinganiso ze-retrieval (Rag). Uhlelo lokusebenza lwakhelwe ngokuphephile futhi lusebenzisa izinsizakalo ze-AWS ukufeza lokhu. Isevisi yokulawulwa kwe-AWS Key (ama-AWS KMS) yenza lula inqubo yokubethela idatha ngaphakathi kohlelo lokusebenza, kanye ne-Amazon API Gateway isekela izinhlelo zokusebenza zokuthola izindawo zokugcina. Ukuqinisekiswa kokuqinisekiswa komsebenzisi nokugunyazwa kusekelwa i-Amazon Cognito, enikeza ubunikazi beKhasimende eliphephile nelikalibekile kanye nokuphathwa kwamakhono (CIAM).

Uhlelo lokusebenza lusebenza kwingqalasizinda ye-AWS lusebenzisa izinsiza ezenzelwe ukuvikela futhi lube sculoble njenge-Amazon S3, AWS Lambda, kanye ne-DynamoDB.

Amathiphu nezincomo

Ukwakha ikhambi lomsizi we-AI lababaleki abasebenzisa i-Amazon Bedrock namanye ama-AWS Services anikeze imininingwane ebalulekile ekwakheni izixazululo ezinomthelela we-AI-Powered Anivenitianed. Ngalesi sivumelwano, sathola ukucatshangelwa okusemqoka ukuthi izinhlangano kufanele zikukhumbule lapho zithuthukisa izixazululo ezifanayo. Isipiliyoni siqokomise ukubaluleka kokulinganisa amakhono obuchwepheshe nge-Hunti-Centric design, ukuhlinzeka ngokusekelwa kwezilimi eziningi, ukulondolozela ubumfihlo bedatha, kanye nokwakha izixazululo ezi-scwable. Lokhu kubhekwa kungasebenza njengesisekelo sezinhlangano ezifuna ukusebenzisa ubuchwepheshe be-AI nefu ukuxhasa izimbangela zobuntu, ikakhulukazi ekwakheni usizo lwedijithali olutholakalayo lwabantu abasusiwe. Okulandelayo kuyisisekelo

• Sebenzisa inkundla yokudlala ye-Amazon Bedrock ukuvivinya uhlangothi lwama-LLMS amaningi eceleni ngokusebenzisa ngokushesha okufanayo. Lokhu kukusiza ekutholeni imodeli enikeza ikhwalithi enhle kakhulu, isitayela, nezwi lokuphendula ngempendulo yakho ethile yokusebenzisa (isibonelo, ukunemba okuyiqiniso vs.
• Uhlole ukukhishwa nezilungiselelo ukwenza ngcono izimpendulo.
• Gcina izindleko engqondweni; Setha ukuqapha kanye nezabelomali kuma-AWS.
• Ngemisebenzi ebandakanya ukubuyisa imininingwane noma ukusesha kwe-semantic, khetha imodeli yokushumeka ngenkathi uqinisekisa ukukhetha izilungiselelo ezifanele. Naka usayizi wokushumeka, ngoba ama-veectors amakhulu angathatha incazelo eyengeziwe kodwa angakhulisa izindleko. Futhi, hlola ukuthi imodeli isekela izilimi isicelo sakho esidingayo.
• Uma usebenzisa isisekelo solwazi, sebenzisa i-Amazon Bedrock Informal Base Ingcabana yokudlala ukuze uvivinye indlela yokuqukethwe itholwe nokuthi kubuyiselwa izindinyana eziningi zitholwa ngombuzo ngamunye. Ukuthola inani elifanele lamavesi abuyiselwe angenza umehluko omkhulu ekutheni izimpendulo zokugcina zicacile futhi zigxile kakhulu ezimpendulweni zokugcina – kwesinye isikhathi zimbalwa, kwesinye isikhathi zisebenza kancane, zisebenza kakhulu ama-chunks asebenza kangcono kunokuthumela umongo omningi.
• Ukuphoqelela ukuphepha nobumfihlo, sebenzisa i-Amazon Bedrock Guardrails. Ama-Guardrali angasiza ukuvikela imodeli ukuthi ivuke imininingwane ebucayi, njengedatha yomuntu siqu noma okuqukethwe kwebhizinisi langaphakathi, futhi ungavimba izimpendulo eziyingozi noma usebenzise isitayela esithile.
• Qala nge-prototype elula, vivinya ikhwalithi yokushumeka esizindeni sakho, futhi unwebe nge-iteratively.

Ukuhlanganiswa kanye nongqimba we-ngcono

I-Bevar Ukraine yelule ingqalasizinda yezezimlando ezi-Core AWS ngobuchwepheshe obuningana obuhambisanayo:

• I-Pinecone Vector database – Isitoreji esisebenzayo kanye nokubuyiselwa kwemali yokugodda kwe-semantic
• Uhlaka lwe-DSpy – Okwenzelwe ukuxhaswa okuhlelekile kanye nokwenza kahle kwe-anthropic's Claude 3.5 Izimpendulo zeSonnet
• I-EasyWeek – Ukuhlelwa kokuqokwa kanye nokuphathwa kwezinsizakusebenza
• Telegraph api – ukulethwa kwe-UI
• I-Amazon Bedrock Guardrails – Ukuphoqelela Inqubomgomo Yezokuphepha
• Ukuqwashiswa kwe-Amazon – Ukuqinisekiswa kwedokhumenti
• I-GitHub-based ukuhlanganiswa okuqhubekayo nokulethwa (CI / CD) Pipeline – Ukuhanjiswa kwesici okusheshayo

Ukuqonda okubalulekile kwezobuchwepheshe

Ukuqaliswa kwembula ukucatshangelwa kwezobuchwepheshe eziningana. Uhlaka lwe-DSPY lwalubalulekile ekwakheni nasekuthuthukiseni izimbangela zethu zolimi lwemodeli. Ngokuhlanganisa izingqimba ezengeziwe zokubonisana kanye namathuluzi okuqwashisa umongo, i-DSPY ngcono ukunemba kokuphendula, ukuvumelana, kanye nokujula. Iqembu lithole ukuthi ukuklama isisekelo solwazi oluqinile nge-metadata ephelele kwakuyisisekelo ekusebenzeni kahle kohlelo.

Ukuthobela i-GDPR kudingekile izinqumo zokwakha ngokucophelela, kufaka phakathi ukuncishiswa kwedatha, isitoreji esivikelekile, kanye nezindlela zokuvuma zomsebenzisi ezicacile. Ukusebenziseka kwezindleko kufinyelelwe ngokusetshenziswa kwamasu ezimo ze-EC2 Spot kanye nokuqaliswa kwesicelo se-API Stratting, okuholela ekusebenziseni okubalulekile kokusebenza ngaphandle kokuyekethisa ukusebenza.

Izithuthukisi zesikhathi esizayo

I-portmap yethu yomgwaqo ifaka ukuthuthuka kwemininingwane eminingi yezobuchwepheshe ukuthuthukisa amakhono ohlelo:

• Ukuqalisa ukuhambisa okuthuthukile kokuthumela kusetshenziswa ama-algorithms wokufunda umshini ukwenza ngcono ukuxhumanisa kwensiza kuwo wonke ama-Domain Domain
• Ukuthuthukisa uhlelo lokuqinisekiswa kwabantu oluyinkimbinkimbi lwamacala alunkimbinkimbi adinga ukwengamela uchwepheshe
• Ukufuduka Kwezakhi Ezifanele Ezakhiweni Zobuciko Ezingenasici Besebenzisa I-Lambda ukwenza Ukusetshenziswa Kwezinsiza Nezindleko
• Ukuthuthukisa isisekelo solwazi ngamakhono wokucinga osezingeni eliphezulu we-semantic kanye nokubuyekezwa kokuqukethwe okuzenzakalelayo

Umphumela

Lesi sixazululo, esisebenza amakhulu ababaleki base-Ukraine eDenmark nsuku zonke, sibonisa amandla ezinsizakalo ze-AWS ekwakheni amasistimu aphethwe amandla, aphephile, futhi asebenza kahle. Ngenxa yalokho, amavolontiya kanye nabasebenzi baseBevar Ukraine basindise izinkulungwane zamahora, futhi esikhundleni sokuphendula imibuzo ephindaphindayo kubabaleki, bangabasekela ezimweni eziyinkimbinkimbi zokuphila. Kwababaleki, umsizi weVictor uVictor ukwesekwa kwempilo yonke evumela abasebenzisi ukuthi bathole izimpendulo emibuzweni ecindezela kakhulu mayelana nemizuzwana yomphakathi eDenmark kanye neminye imibuzo ngemizuzwana esikhundleni sokuthola usizo. Njengoba kunikezwe i-GAST Knowledge Base Victor isebenzisa ukukhiqiza izimpendulo, ikhwalithi yokusekelwa seke yathuthuka.

Ukugcina

Ngokuklanywa ngokucophelela ukwakhiwa kwezakhiwo kanye nokuhlanganiswa kobuchwepheshe obuhambisanayo, sidale ipulatifomu ebhekana nempumelelo nezinselelo ababhekana nababaleki lapho bevikelekile kanye nokuvikelwa kwedatha.

Impumelelo yalokhu kuqaliswa ihlinzeka nge-Blueprint Yezixazululo ezifanayo Kwezinye izizinda zenkonzo yezenhlalo, ababaleki abasesisezingeni lomhlaba kanye nabanye abantu abadinga ukuhlanganisa ingqalasizinda yohlelo olunamandla ukudala umthelela wefu ocabangayo ukudala umthelela we-system.

Mayelana nababhali

Taras Tsarenko Imenenja yohlelo eBevar Ukraine. Isikhathi esingaphezu kweshumi emhlabeni wezobuchwepheshe, ama-taras aholele konke kusuka kumaqembu agugile asolwandle angu-5 noma ngaphezulu kwabantu abahlangene, noma ngabe kuwukuphuma kwezinkinga zomsebenzi, noma ukuxazulula izinkinga eziyinkimbinkimbi, noma ukuqinisa amaqembu ukudala imikhiqizo enenjongo. Ama-Taras abhekene nezixazululo eziqhutshwa yi-AI kanye nobunjiniyela bedatha, ubuchwepheshe obunamandla njengokufunda komshini kanye ne-ai ekhiqizayo usebenzisa i-Amazon Sagemaker Ai, i-Amazon Bedrock, I-Amazon OpenSense Search, nokuningi. I-Taras iyi-AWS eqinisekisiwe ye-ML Jioner Associate.

U-Anton Garvanko Ingabe uchwepheshe wokuthengisa wokuthengisa wokuhlaziya we-Europe enyakatho. Njengomthengisi wezezimali waphenduka umthengisi, u-Anton wachitha iminyaka eyi-15 ezindimeni ezihlukene zobuholi zezimali ezingeni lokuhlinzekwa kanye nezimboni zezezimali kanye nezimboni zezezimali. U-Anton wajoyina ama-Amazon eminyakeni emihlanu edlule futhi ubeyingxenye yamaqembu okuthengisa ochwepheshe agxile kubuhlakani bebhizinisi, ama-analytics, kanye ne-ai ekhiqizayo iminyaka engaphezu kwemithathu. Unothando ngokuxhuma umhlaba wezimali kanye nokwenza isiqiniseko sokuthi ubuhlakani bebhizinisi kanye nokuhlaziya okunikwa amandla ngokuxhaswa kwe-AI verationess Airnies nsuku zonke izimboni futhi basebenzise amacala.

Vitalii Bozadzhy Ungunjiniyela ophezulu onolwazi olunzulu ekwakheni izisombululo eziphezulu, izixazululo ezisuselwa efwini, ochwepheshe eJava, e-golang, esheshayo, nasePython. Usebenza ngokukhethekile ezinhlelweni ze-bakendles ezihlelekile, izakhiwo ze-Microservace zenzelwe ukushintshanisa nezinqubo zebhizinisi, kanye nokwakha izingqalasizinda zefu ezithembekile nezivikelekile. Ngaphezu kwalokho, unolwazi ekwenzeni ngcono izinsizakusebenza ze-compute futhi wakhe izixazululo ezithuthukile ezihlanganiswe nemikhiqizo. Ubuchwepheshe bakhe buhlanganisa umjikelezo wokuthuthuka okugcwele – kusuka ekuklameni nasekwakheni ukwakhiwa kokuthumela kanye nokugcinwa – ngokugxila okuqinile ekusebenzeni, ngephutha ukubekezelelana, kanye nokwenza izinto ezintsha.

Vladyslav horbatenko Ungumfundi wesayensi yekhompyutha, uSolwazi Assistant, kanye nososayensi wedatha ngokugxila okuqinile kubuhlakani bokufakelwa. UVladyslav waqala uhambo lwakhe ngokufunda ngomshini, nokuqiniswa okujulile, futhi kancane kancane waba nesifiso esikhulu ngamamodeli amakhulu wezilimi (LLMS) nomthelela wawo. Lokhu kwaholela ekutheni akhulise ukuqonda kwakhe ama-llms, futhi manje asebenza ekuthuthukiseni, egcina, futhi ethuthukisa izixazululo ezisuselwa kwi-LLM. Unikela ngamaphrojekthi amasha ngenkathi uhlala usesikhathini nentuthuko yakamuva e-AI.