Ungawakha Kanjani Ama-Agent E-AI Anamandla Ngamathuluzi ku-Csharp

Ikhaya » Python » Ungawakha Kanjani Ama-Agent E-AI Anamandla Ngamathuluzi ku-Csharp

Ukwakha ama-ejenti e-AI ngamathuluzi ku-C# kushintshe kusuka ekuhlolweni kocwaningo kuya endleleni ewusizo kakhulu yokwengeza ubuhlakani bangempela kuzinhlelo zokusebenza zebhizinisi. Izinhlaka zesimanje ezivela ku-Microsoft kanye nama-SDK e-OpenAI ne-Azure OpenAI zakamuva zenza kube nokwenzeka ukudlulela ngale kwama-chatbot alula, ukuxhuma amamodeli ezilimi ezinkulu nekhodi, amafayela, imisebenzi yokusebenza kanye nezinhlelo zebhizinisi ngenkathi kusalawulwa ukuphepha, izindleko kanye nokuthembeka.

Lo mhlahlandlela ukuqondisa ngemiqondo eyinhloko, izinqumo zokwakha kanye nezibonelo ze-.NET ozidingayo ukuze uklame ama-ejenti alungele ukukhiqiza ku-C#. Sizohlanganisa imibono evela ku-Azure OpenAI Assistants, i-Microsoft Agent Framework, amaphethini okuhlelwa, ukuhlolwa, ukubonwa kanye nokusetshenziswa kwamafu, sichaze ukuthi konke kuhambisana kanjani nesu elihlangene lezinhlelo zokusebenza zangempela.

Ukuthi i-ejenti ye-AI iyini ngempela (nokuthi kungani ibalulekile ku-.NET)

Kuhlelo lwe-.NET, i-ejenti ye-AI iqondwa kangcono njengengxenye yesofthiwe eqhutshwa umgomo enikwe amandla yi-LLM engakwazi ukucabanga, ukukhetha amathuluzi nokusebenza ngaphakathi kohlelo lwakho lokusebenza. Esikhundleni seskripthi esiqinile esihlala silandela indlela efanayo, i-ejenti yamukela okokufaka okuvulekile, inqume ukuthi yini ezoyenza ngokulandelayo bese isebenzisa ikhodi yakho nedatha ukuze ifinyelele emphumeleni.

Ama-ejenti aba usizo kakhulu uma ungeza amakhono amathathu phezu kokukhiqiza umbhalo ocacile. Ubanika ukucabanga nokwenza izinqumo (ngokusebenzisa ama-LLM, usesho noma ama-algorithms okuhlela), ikhono lokubiza amathuluzi (imisebenzi yendawo ye-C#, amaseva e-MCP, ama-API, ukwenziwa kwekhodi) kanye nokuqwashisa ngomongo (umlando wengxoxo, izintambo, izitolo ze-vector, amagrafu olwazi lwebhizinisi noma usesho lwamafayela). Yilokhu okuguqula ukuqedwa kwengxoxo elula kube yingxenye engaxhumanisa umsebenzi wezinyathelo eziningi ngokuzimela.

Njengoba imigomo yakho iba yinkimbinkimbi kakhulu, awuvamile ukusebenzisa yonke into njengesikhuthazo esisodwa esikhulu esingacacile; uhlukanisa umsebenzi ube yimisebenzi yokusebenza. Ukuhamba komsebenzi kuwuchungechunge noma igrafu yezinyathelo ezidingekayo ukuze kufinyelelwe umgomo: ukuqoqa izidingo, ukuklama, ukusebenzisa, ukuhlola nokufaka isici, isibonelo. Isinyathelo ngasinye singaqukatha imisebenzi engaphansi futhi singase sibuyele emuva kuye ngamaphutha noma ulwazi olusha, ngakho-ke ukuhlela ngokushesha kuba yinto ekhathazayo yekilasi lokuqala.

Uma ubeka ama-ejenti ngaphakathi kwalezi zindlela zokusebenza, uthola ukugeleza komsebenzi we-agent: ukugeleza lapho amanxusa ebambisana khona ukuze enze, avumelanise futhi athuthukise imisebenzi. Ungase ube ne-ejenti ehlaziya ama-log, enye ebhala ukulungiswa kwamakhodi, kanti eyesithathu elungiselela imibiko yabathintekayo. Ingxenye ebalulekile yindlela abadlulisa ngayo ulwazi, indlela abahlanganiswa ngayo nokuthi ugcina kanjani lonke uhlelo lubonakala futhi luhloleka.

Izakhi eziyinhloko zabasizi be-AI kanye nama-ejenti

Amapulatifomu amaningi e-AI esimanje ahloselwe i-C# kanye ne-.NET abelana ngesethi encane yezingxenye eziyinhloko, noma ngabe ukuqanjwa kwegama kuhluka kancane phakathi kwe-Azure OpenAI Assistants kanye ne-Microsoft Agent Framework. Ukuqonda la mabhulokhi kukusiza ukuthi uklame ukwakheka kwakho esikhundleni sokukopisha izingcezu ngokungazi.

Umsizi noma i-ejenti yiklayenti eliyinhloko le-AI elisebenzisa ukucushwa kwe-LLM plus ukucubungula imiyalelo, ukuphatha izingxoxo kanye nokubiza amathuluzi. Ku-Azure OpenAI Assistants, le nto igoqa ukucushwa kwemodeli, imiyalelo, kanye nokucushwa kwamathuluzi. Ku-Microsoft Agent Framework, i- AIAgent ihlanganisa iklayenti lengxoxo (i-OpenAI noma i-Azure OpenAI) kanye namathuluzi nemiyalelo, futhi ayinabo ubunikazi ngamabomu ukuze ikwazi ukusebenzela izingxoxo eziningi ngasikhathi sinye.

Ingxoxo noma iseshini imelela ingxoxo eyodwa phakathi komsebenzisi kanye ne-ejenti, okuhlanganisa yonke imiyalezo kanye nesimo esifanele. Abasizi be-Azure OpenAI bakhuluma ngakho izindikimba, eziphethe imiyalezo futhi zisingatha ukunqunywa okuzenzakalelayo ukuze kufanelane nomongo wemodeli. I-Microsoft Agent Framework ikhuluma nge I-AgentSession, equkethe umlando futhi ingahlelwa ngokulandelana futhi igcinwe. Zombili zifeza injongo efanayo: ukulandelela umongo phakathi kwamajika amaningi.

Imilayezo iminikelo yomuntu ngamunye ngaphakathi kwengxoxo noma iseshini, ekhiqizwa abasebenzisi noma umsizi. Imilayezo ingaqukatha umbhalo ocacile, izithombe noma amafayela, futhi kuma-API Omsizi agcinwa njengohlu olu-odiwe ngaphakathi kwentambo. Ohlangothini lwe-C#, uvame ukuwathola njengamaqoqo athayishwe ngokuqinile lapho ungahlola khona umbhalo, izichasiselo kanye nezinkomba zamafayela.

Ukuqalisa, ukwenza noma ukunxusa kuyindlela eyodwa yokwenza i-ejenti isebenze phezu kwengxoxo noma iseshini ethile. Uthatha umongo okhona, uwuthumele kumodeli kanye namathuluzi nokucushwa, bese ulinda kuze kube yilapho ukugijima kufinyelela esimweni sokugcina. Ngesikhathi sokugijima, i-ejenti ingakhiqiza imiyalezo emisha, ishayele amathuluzi futhi ibuyekeze intambo noma isimo seseshini.

Izinyathelo zokwenza zakha umkhondo onemininingwane wakho konke okwenzekile ngesikhathi sokusebenza kwe-ejenti. Umsizi angashayela ithuluzi lokusesha ifayela, aqalise umhumushi wekhodi, noma abize umsebenzi owenziwe ngokwezifiso izikhathi eziningi njengoba echaza ngomsebenzi. Ukuba nombono ohlelekile walezi zinyathelo kuyasiza kakhulu ukuqonda ukuthi kungani impendulo ethile yakhiqizwa kanye nokulungisa noma ukuhlola ukuziphatha kamuva.

Ukudala i-ejenti encane yekhonsoli ye-C# nge-Azure OpenAI Assistants

Ukuze ubone le mibono isebenza, ungavula uhlelo lokusebenza lwekhonsoli elincane le-.NET elisebenzisa ama-SDK asemthethweni e-OpenAI noma i-Azure OpenAI ukwakha umsizi ofunda idatha evela kumafayela futhi akhiqize ukubonakala. Umqondo uwukuxhumanisa i-LLM kokubili ekusesheni amafayela nasekusetshenzisweni kwekhodi, bese uyivumela iphendule imibuzo yokuhlaziya ngolimi lwemvelo.

Isinyathelo sokuqala ukusetha iphrojekthi: dala uhlelo lokusebenza olusha lwekhonsoli ye-.NET bese wengeza amaphakheji e-NuGet e-OpenAI kanye ne-Azure.AI.OpenAI. Bese ufaka amakhasimende ayinhloko ku Program.cs, kungaba nge-OpenAI ngqo noma nge-Azure OpenAI kusetshenziswa isitifiketi esifana ne- DefaultAzureCredentialKuklayenti le-OpenAI uthola i- AssistantClient ukuphatha abasizi kanye neqembu elihlukile OpenAIFileClient ukuze kulayishwe amafayela.

Okulandelayo ulungiselela idatha engokoqobo ukuze umenzeli asebenze ngayo ngokwakha idokhumenti enkumbulo, uyihlele njenge-JSON bese uyithumela kuklayenti lefayela. Esibonelweni, le JSON ifaka ikhodi yezinyanga eziningana zokuthengiswa komkhiqizo kwenkampani eqanjiwe, ihlanganisa izinyanga ngobuningi bomkhiqizo ngamunye. Ngokuyilayisha nge Assistants ngenhloso yefayela, ulimaka njengento engaseshwa yi-ejenti.

Uma idatha isikhona ohlelweni, ulungiselela umsizi nge- AssistantCreationOptions ukuze unike amandla kokubili ukusesha amafayela kanye nethuluzi lokuhumusha ikhodi. Ucacisa igama, isethi yemiyalelo ecacile (“ungumsizi obheka idatha yokuthengisa futhi akhiqize imiboniso lapho ebuzwa”), bese unamathisela amathuluzi: a FileSearchToolDefinition ukuze umsizi akwazi ukubuza amafayela, kanye ne- CodeInterpreterToolDefinition ukuze ikwazi ukubhala nokusebenzisa ikhodi endaweni ene-sandbox ukuze ihlaziywe noma ikhiqizwe ishadi.

Ukuze usesho lwefayela lusebenzise idokhumenti yakho yokuthengisa elayishiwe, uyihlobanisa nesitolo esisha se-vector ngaphakathi ToolResources. Umsizi VectorStoreCreationHelper ibopha i-ID yefayela elilayishiwe esitolo se-vector umsizi angasibuza ngokwesisho esikhundleni sokuskena umbhalo ongahluziwe. Lena indlela elula kodwa enamandla yokwengeza ukuziphatha kokuthola okuthuthukisiwe.

Uma unezinketho ezikhona, udala umsizi ngokudlulisa imodeli eqondiwe (isibonelo gpt-4o) kanye nokucushwa, bese uphendula ingxoxo enomyalezo wokuqala womsebenzisi. Leso sicelo sokuqala singaba yinto efana nokuthi “Umkhiqizo i-113045 usebenze kanjani ngoFebhuwari? Hlela ukuthambekela kwawo ngokuhamba kwesikhathi.” Ekugcineni, ushaya ucingo. CreateThreadAndRun, okudala intambo futhi kuqalise ukugijima.

Ngenxa yokuthi ukugijima akuhambelani ngokwemvelo, uhlelo lokusebenza lwekhonsoli luvame ukukhetha ukugijima kuze kube yilapho isimo siba yisiphetho. Ngemva kwalokho, udonsa imiyalezo yemicu ngokulandelana okukhuphukayo bese uyiphinda-phinda: ukuphrinta umbhalo womsizi, ukukhipha izichasiselo zemibhalo noma amafayela akhiqizwe, bese ulanda imiphumela yesithombe usebenzisa iklayenti lefayela ukuze ukwazi ukulondoloza amashadi akhiqizwe ngumhumushi wekhodi kudiski njengamafayela e-PNG.

Umphumela wokugcina uba uhlelo lokusebenza lwe-C# console oluzimele lapho umsizi oyedwa angasesha khona idatha yokuthengisa ehlelekile, enze izibalo ngekhodi futhi abuyisele kokubili ukuqonda kombhalo kanye namagrafu abonakalayo ku-loop ezenzakalelayo ngokuphelele. Le phethini ikhula kahle ibe yi-backends yewebhu noma izinsizakalo zangemuva uma ungeza ukuphikelela nokuqinisekiswa.

Ukuklama ukwakheka kwe-ejenti okuqinile ku-C#

Uma ushintsha kusuka ku-demo uye kuhlelo lokusebenza lwangempela, indlela ohlela ngayo ama-ejenti akho ibaluleke kakhulu njengokuthi ukhetha yiphi imodeli. Ukwakhiwa okuhle kwenza kube lula ukuhlola, ukukala, ukuvikela nokuthuthukisa ikhambi lakho ngaphandle kokugcina unezinkinga eziningi zokucela kanye nokuphinda ubize.

Isu eliqinisekisiwe ukuphatha izinto ezibangela lokhu njengezakhi ezikhethekile kunokuphatha ubuchopho obubodwa "obungenza konke". Isibonelo, ungase uchaze i-ejenti eyodwa egxile ekutholeni nasekuqinisekiseni ulwazi, enye i-ejenti ezinikele ekubhaleni nasekufingqeni okuqukethwe, kanye nenye enomsebenzi wayo kuphela ukusebenzisana nama-API noma izizindalwazi zangaphandle. Lokhu kuhlukaniswa kuvumela ukuhlolwa kweyunithi okuqondiwe, ukuthunyelwa okuzimele kanye nemikhawulo yokuphepha ecacile kanye nethokheni.

Isimo kanye nenkumbulo kuba yizithiyo ngokushesha uma uzibheka njengento ecatshangelwe kamuva. Imlando yengxoxo ikhula ngokuhamba kwesikhathi, futhi ukuthumela umbhalo wonke ngokungazi kumodeli kuyo yonke indlela kwandisa ukubambezeleka kanye nezindleko. Amasu asebenzayo afaka phakathi ukufingqwa kwemiyalezo yangaphambilini ngezikhathi ezithile, ukuhlukanisa izingxoxo zibe imicu ehlukene ngomsebenzisi ngamunye noma ngokwesimo ngasinye sokusetshenziswa, kanye nokusebenzisa izinqubomgomo zokuhlanganisa ezisekelwe ekubalulekeni kwencazelo ukuze izingxenye ezibalulekile kakhulu zesikhathi esidlule zigcinwe ngokuningiliziwe.

Ezimweni zokukhiqiza ufuna futhi isitolo esiqhubekayo senkumbulo ukuze izingxoxo zikwazi ukusinda lapho inqubo iqala kabusha, ukwehluleka noma ukuthunyelwa kabusha. Izinhlaka zama-ejenti ezifana ne-Microsoft Agent Framework zenza amaseshini alandelelwe ku-a JsonElement, ongayifaka ku-SQL Server, Redis noma kunoma yisiphi isitolo se-NoSQL. Lelo khono elifanayo livumela izindlela zokuhlola kanye nokuhambisana nemithetho ngoba ungakha kabusha isimo i-ejenti eyayinaso lapho yenza isinqumo.

Amathuluzi kanye nezingcingo zomsebenzi yilapho amanxusa eyeka khona ukungenzi lutho bese eqala ukwenza umsebenzi owusizo. Ukuveza izindlela zendabuko ze-C# njengamathuluzi kuvumela imodeli ukuthi isebenzise ukuziphatha okufana nokubuza i-CRM, ukusebenzisa ukuhlaziya idatha noma ukuqala imisebenzi yokusebenza. Ithuluzi ngalinye kufanele libhalwe nge-metadata ecacile (izincazelo kanye nemibhalo yamapharamitha), ukuze i-LLM yazi ukuthi kufanele ilibize nini nokuthi yiziphi izimpikiswano.

Ngenxa yokuthi ithuluzi lokuziphatha kabi lingaphula ukusebenzisana konke, udinga ubunjiniyela obuqinile obuzungezile: ukuqinisekiswa kokufaka, isikhathi sokuvala, ukuphathwa kwe-exception kanye ne-guardrails. Ungacabangi ukuthi imodeli ihlala idlula izimpikiswano eziphelele; qinisekisa amapharamitha futhi uhlanze noma yiziphi izingcingo zangaphandle. Cabanga futhi ngokwezilinganiso kanye nemikhawulo yamanani ngethuluzi ngalinye ukuze ugweme izindleko ezisheshayo noma ukugcwala ngokweqile kwezinhlelo ezingezansi ngengozi.

Ezimweni ezikhangayo, ukuhlelwa kwama-ejenti amaningi kungavula amakhono okunzima ukuwafeza nge-ejenti eyodwa eyodwa. Ungaxhumanisa i-ejenti "yomcwaningi" eqoqa futhi ihlole ulwazi, "umhlaziyi" ochaza okutholakele, kanye "nombhali" okuguqula kube yimibiko, ngayinye ixhumana ngemiyalezo ehlelekile futhi yabelana ngomsebenzi (njengedokhumenti ehlanganyelwe noma isitolo solwazi). Le ndlela ikhulisa ubuchwepheshe futhi yenza indlela yokwenza izinqumo ilandeleke uma kamuva udinga ukubuyekeza noma ukuhlola imiphumela.

Kusukela ku-Semantic Kernel kanye ne-AutoGen kuya ku-Microsoft Agent Framework

I-Microsoft ibilokhu ihlanganisa amathuluzi ayo e-ejenti ye-.NET, ihlanganisa imibono evela ku-Semantic Kernel kanye nephrojekthi ye-AutoGen ibe yi-Microsoft Agent Framework entsha, ehlangene (i-MAF). Lolu hlaka luhlose ukukunikeza ukuzinza kwezinga lebhizinisi kanye nezici ngenkathi lwenza kube lula indlela owakha ngayo ama-ejenti amaningi kanye nemisebenzi yokusebenza esekelwe kugrafu.

I-MAF okwamanje itholakala emphakathini futhi iyatholakala kokubili ku-.NET kanye ne-Python ngaphansi kwelayisensi ye-MIT. Ngisho noma amanye ama-API esashintsha phakathi kwabantu abazokhishwa, isiqondiso siphelele sicacile: Ama-AIagents okuziphatha okuhlakaniphile, ama-AgentSessions okuphathwa kwesifundazwe, kanye nohlelo lokusebenza olusekelwe kumagrafu kanye nabaphathi bezindlela zokulawula eziqondile.

Empeleni, uhlaka luhlukanisa phakathi kwama-ejenti kanye nemisebenzi yokusebenza, ngayinye ehloselwe izimo zezinkinga ezahlukene. Ama-ejenti ayizinhlelo ezishintshashintshayo ezisebenzisa ama-LLM ukuhumusha okokufaka, ukunquma ukuthi yimaphi amathuluzi okufanele awabize futhi akhiqize izimpendulo. Akhanya ezizindeni ezingalindelekile njengezingxoxo zokusekela ubuchwepheshe lapho abasebenzisi bangabuza noma yini. Ukugeleza komsebenzi, ngokuphambene, kuwukulandelana kwezinyathelo okucacile okuxhunywe njengamagrafu futhi kusetshenziswa uma ufuna ukucubungula okuqondile, okuchazwe kahle njengemibhobho yedatha noma izintambo zokugunyaza.

Isiqondiso esisemthethweni singafingqwa ngokuthi “uma ungakwazi ukusebenzisa umsebenzi njengomsebenzi ojwayelekile, cishe awudingi i-ejenti yawo.” Ngamanye amazwi, gcina ama-ejenti ezizindeni lapho ungakwazi ngempela ukuchaza zonke izinyathelo kusengaphambili, bese uthembela emisebenzini noma ikhodi yakudala ukuze uthole ukugeleza okuphindaphindwayo nokuqinisekile. Ukuxuba kokubili ezindaweni ezifanele kuyisihluthulelo sokwakha izinhlelo ezilondolozekayo.

Ukuze wenze lokhu okuqondile, cabanga nge-chatbot yokusekela eyakhiwe njenge-ASP.NET Core 10 API kusetshenziswa i-Microsoft Agent Framework. I-ejenti isebenzisa iklayenti lengxoxo (elisekelwa yi-Azure OpenAI noma i-OpenAI) njengenjini yayo yokucabanga, futhi inhloso yayo eyinhloko ukuphendula imibuzo mayelana nemibhalo yangaphakathi egcinwe kumafayela e-Markdown ngenkathi igcina umongo kuyo yonke imiyalezo eminingi evela kumsebenzisi ofanayo.

Ngokuthakazelisayo, isibonelo singayeqa ngamabomu i-RAG ngokufaka ama-embeddings futhi sihlale singokoqobo ngokusebenzisa ukusesha amagama angukhiye ngaphezu kwamafayela ayisicaba njengendawo yokuqala. Lokho kugcina ukugxila endleleni i-MAF eyakha ngayo i-ejenti, amathuluzi kanye nezikhathi esikhundleni sokulahleka ekucushweni kwesizindalwazi se-vector, kuyilapho isasekela ukusebenzisana kokusekela okunokwenzeka kakhulu.

Imiqondo emihlanu ebalulekile ku-Microsoft Agent Framework

Izifundo ezisemthethweni ze-MAF zihlela ukufunda ngemibono emihlanu eqhubekayo ehambisana kahle nendlela abathuthukisi be-C# asebecabanga ngayo ngezinsizakalo kanye nesimo. Ukuzivumelanisa nezimo ngale mibono kukunikeza isisekelo esiqinile sanoma yimuphi umenzeli ozokwakhela kuye i-.NET.

Okokuqala kuza i-ejenti yakho yokuqala: i AIAgent yakhiwe ngeklayenti lengxoxo, imiyalelo kanye negama. Ukhomba umenzeli kumodeli yengxoxo enikezwa yi-AzureOpenAIClient noma i-OpenAI, ohlinzeka ngesiqondiso sezinga lesistimu (“ungumsizi wokusekela owusizo”) bese ushayela ucingo. RunAsync ngokufakwa komsebenzisi. Imininingwane ebalulekile ukuthi isibonelo se-ejenti asinasimo futhi singakhonza izingxoxo eziningi ezizimele ngesikhathi esisodwa.

Okwesibili amathuluzi, okuyizindlela ze-C# ezihlotshiswe ngazo izimfanelo futhi ziguqulwe zibe imisebenzi engashintshwa ngokusebenzisa AIFunctionFactory.Create(). Uma i-ejenti isebenza, i-LLM ithola i-schema ethathwe kulezo zimfanelo futhi inganquma ngokuzimela ukuthi izolibiza nini futhi kanjani ithuluzi ngalinye, kufaka phakathi izimpikiswano. Yilapho i-logic yebhizinisi lakho kanye nokuhlanganiswa kwangaphandle kuba yingxenye yesikhala sesenzo se-ejenti.

Okwesithathu ukusekelwa kwengxoxo ephendula ama-multi-turn, okwenziwa yi-MAF AgentSession izinto. Ngoba AIAgent ngokwayo ayikhumbuli lutho, ingxoxo ngayinye eqhubekayo ihlala ngaphakathi kweseshini eyenziwe nge CreateSessionAsync()Udlulisa leso sikhathi emuva ezingcingweni ezilandelayo, uvumela umenzeli ukuthi alandelele imiyalezo yangaphambilini, izintandokazi zomsebenzisi kanye nezinkinga ezingaxazululwanga.

Okwesine inkumbulo nokuphikelela, okuvunyelwe yiqiniso lokuthi amaseshini angahlelwa ngokulandelana abe yi- JsonElement. Lokho kwenza kube lula ukuzigcina kwimemori, iRedis, ithebula le-SQL noma kunoma yisiphi esinye isitolo osithandayo, bese uziphinda uzisebenzise kabusha DeserializeSessionAsync(). Ngezimo zokusekela, lokhu kusho ukuthi umsebenzisi angavala isiphequluli sakhe bese kamuva aqalise ingxoxo efanayo, noma isevisi ehlukile ingathatha izintambo kalula ngemva kokuqala kabusha.

Okwesihlanu yimisebenzi yokusebenza, eyakhiwe nge WorkflowBuilder uma udinga ukuhlela ngokusobala ama-ejenti amaningi noma izinyathelo zokucubungula ezilandelanayo. Uchaza ama-executors njengamayunithi okucubungula, uwaxhumanise ngama-edges bese uvumela injini yomsebenzi ukuthi iphathe umzila kanye noshintsho. Ezimweni eziningi zengxoxo ngeke udinge umzila wokusebenza nhlobo, kodwa uba usizo kakhulu uma ufuna umzila ohleliwe, ukuhlukaniswa noma izinyathelo zomuntu ngaphakathi kwe-loop ezizungeze ama-ejenti akho.

Ukusebenzisa i-bot yokusekela yangempela nge-MAF, amathuluzi kanye nezikhathi

Isampula eqondile ekhombisa imiqondo engenhla yi-SupportBot API esekelwa yiphrojekthi ye-ASP.NET Core 10. Le nsizakalo idalula indawo yokugcina ye-HTTP eyamukela imiyalezo yomsebenzisi kanye nesihlonzi seseshini, idlulisele izizathu ku-AIAgent futhi iqhubeke neseshini ukuze umongo ulondolozwe kuzo zonke izicelo.

Ithuluzi eliyinhloko kulesi simo yiDocumentationTool eyaziyo indlela yokusesha amafayela angaphakathi eMarkdown. Umthwalo wemfanelo wayo ukuthola iziqondiso ezifanele, ama-FAQ noma ama-module manual kanye nokubuyisela izingxenye zombhalo ezisiza umenzeli ukuthi abhale impendulo. Izimfanelo ezisetshenziswa ezindleleni zayo azizona ezokuhlobisa; i-MAF izisebenzisa ukwakha uhlelo lomsebenzi olufundwa yi-LLM, futhi ukucaca kwalezo zincazelo kuthonya kakhulu ukuthi imodeli ikhetha futhi ibize ithuluzi ngempumelelo kangakanani.

Inketho yokuklama esebenzayo ngaphakathi kwaleli thuluzi ukubuyela ekubuyiseleni wonke amadokhumenti uma kungekho okuhambisana nesihloko esiceliwe kahle. Kunokushiya i-ejenti ingenazo nhlobo izinto, kungcono unikeze umongo omningi kakhulu bese uvumela imodeli ukuthi ikhethe izingcezu ezinhle kakhulu kunokuyivumela ukuthi ibone izinto ngendlela engaqondakali. Le ndlela "yokubuyela emuva ephephile" ibonakala kaningi ekusetshenzisweni kwe-ejenti okuqinile.

I-SupportAgentFactory ibe isihlanganisa konke ndawonye ngokuthatha i- AzureOpenAIClient, ukukhipha iklayenti lengxoxo nge GetChatClient(), ukuyivumelanisa ne AsIChatClient() bese uyiguqula ibe AIAgent nge AsAIAgent(). Phakathi nalesi sinyathelo sokugcina, amathuluzi abhalisiwe kanye nemiyalelo iba yingxenye yokucushwa kwe-ejenti esetshenziswa kuyo yonke ingxoxo. Ngokuvamile ubhalisa le ejenti eyakhiwe njenge-singleton esitsheni se-DI ukuze ikwazi ukukhonza izikhathi eziningi ngesikhathi esisodwa.

Ukuphathwa kweseshini kufinyelelwe ngemuva kwe- InMemorySessionStore ngesikhathi sokuthuthukiswa, obamba imihlangano njenge JsonElement amanani. Iphephile ngentambo ConcurrentDictionary kwanele lapha ukugwema ukukhiya ngesandla. Ekusetshenzisweni kwangempela ubungashintsha lokhu kusetshenziswa ngesitolo esisekelwe yiRedis noma esisekelwe yidathabheyisi, ugcine isikhombimsebenzisi siqinile kodwa sithola isitoreji esiqinile kanye nokukhula okuvundlile.

Ubuso be-API ku Program.cs kugcinwa kulula ngamabomu: i-POST eyodwa /chat iphuzu lokugcina elamukela i-ID yeseshini kanye nomyalezo womsebenzisi. Umphathi wesicelo ulayisha noma udala iseshini, asebenzise i-ejenti, ahlele iseshini ebuyekeziwe ngokulandelana ngendlela engavumelanisiwe (qaphela ukuthi SerializeSessionAsync i-async ku-RC1, noma ngabe amadokhumenti okuqala asikisele okuhlukile), iyaqhubeka futhi ibuyisele impendulo yomsizi kuklayenti. Ngokombono we-frontend, "ukuhlala engxoxweni efanayo" kusho ukuthumela i-ID yeseshini efanayo kukholi ngayinye.

Uma usebenzisa i-API futhi uxoxa ngokumelene nayo, ungabuka i-ejenti iphatha umongo phakathi kokushintshana njengommeleli wokusekela abantu. Umlayezo wokuqala ungase uchaze inkinga yokungena ngemvume; umbuzo wesibili, othunyelwe nge-ID yeseshini efanayo, ungabhekisela ku-"lelo phutha futhi" ngaphandle kokuphinda usho imininingwane egcwele, futhi i-ejenti isaphendula ngendlela ehambisanayo ngoba isimo sixhumene nesitolo seseshini.

Ukugeleza komsebenzi kuzoqala ukuthola imali yakho kuphela uma ungeza izici ezifana nokuhlukaniswa kwezinhloso okuzenzakalelayo, ukuhambisa kuma-ejenti akhethekile (ukukhokhisa, ukufinyelela, ukubika) noma ukukhushulelwa kubasebenzi babantu. Ungabe usufaka umphathi wezigaba ngaphambili kwegrafu yomsebenzi bese uyixhuma kuma-ejenti athile ngesihloko, noma wengeze i-node yomuntu ngaphakathi kwe-loop evimba ukwenza okuzenzakalelayo kanye nomongo wezandla kumuntu lapho ukuzethemba kuphansi.

Ukugeleza komsebenzi, izindlela zokuhlanganisa kanye nokusebenzisana kwama-ejenti amaningi

Ngisho nangaphandle kwe-MAF, kuyasiza ukucabanga ngendlela imisebenzi equkethe ama-ejenti ehlelwa ngayo, ngoba isakhiwo sawo sithinta ukubambezeleka, izindleko kanye nokulandelelwa. Kunezinhlobo eziningana zezindlela ezivamile ezibonakala kuwo wonke amaphrojekthi kanye nezinhlaka.

Ukuhlelwa ngokulandelana kusho ukuthi ama-ejenti aphatha imisebenzi ngokulandelana, edlulisela imiphumela phambili. Isibonelo, i-ejenti yokubuyisa idatha iqala ngokuqoqa amadokhumenti afanele, bese iwadlulisela ku-ejenti yokuhlaziya, yona enikeza okutholakele kwayo ku-ejenti yokubika. Lokhu kulula ukukucabanga futhi kulula ukukulungisa, ngezindleko zokubambezeleka okuphezulu kusukela ekuqaleni kuze kube sekupheleni.

Ukuhlelwa ngasikhathi sinye kusebenzisa ama-ejenti amaningi ngesikhathi esisodwa, ngalinye ligxile esicini esihlukile senkinga. I-ejenti eyodwa ingase ibale ama-metric, enye ingase ifune izehlakalo zakamuva, kanti eyesithathu ingase ihlole umthelela wokuthobela imithetho, konke ngesikhathi esisodwa. Uma sebeqedile, umxhumanisi uhlanganisa imiphumela yabo ibe yimpendulo eyodwa. Le ndlela inciphisa ukubambezeleka kodwa idinga ukulawulwa kwezinsiza ngokucophelela kanye nokuxazulula izingxabano.

Ukugeleza kokudluliselwa kushintsha ngokusobala ubunikazi bomsebenzi kusuka komunye umenzeli kuya komunye ngokusekelwe ezimweni noma emiphumeleni ephakathi. Uma i-ejenti yokusekela ithola ukuthi umbuzo uhlobene nokuthengisa, ingadlulisela ingxoxo ku-ejenti yokuthengisa ekhethekile, igcine umlando wengxoxo kanye ne-metadata ngokuzithandela. Lokhu kuyasiza kakhulu ohambweni oluyinkimbinkimbi lwamakhasimende lapho umthwalo wemfanelo uhamba khona phakathi kwamaqembu ngokusemthethweni.

Ukusethwa kwesitayela sengxoxo yeqembu kuvumela amanxusa amaningana ukuthi abambisane esiteshini sengxoxo esabiwe, beshintshana imiyalezo ngesikhathi sangempela. I-ejenti ngayinye iletha umbono wayo noma isethi yamathuluzi, futhi i-orchestrator ephakathi noma umlawuli we-LLM angaphatha ingxoxo ukuze ihlangane kunokuba iqhubeke unomphela. Le ndlela inamandla kodwa idinga izivikelo eziqinile ukuze kugwenywe umsindo nezindleko ezingadingekile.

Ekugcineni, ukuhlelwa kwamagnetic kubeka "umholi" oyedwa noma i-ejenti yomqhubi we-automatike ekuqondiseni abanye. I-ejenti eholayo ihlukanisa umsebenzi, ithumele imisebenzi engaphansi kochwepheshe abafanele bese ihlanganisa imiphumela yayo. Lokhu kufana nomphathi wobunjiniyela ohlanganisa ithimba labathuthukisi futhi kungaveza ukugeleza okucacile nokungahlolwa ezizindeni eziyinkimbinkimbi.

Ukuhlola, ukubonwa, ukulawulwa kwezindleko kanye nokuphepha

Ukuthumela ama-ejenti e-AI ekukhiqizweni ngaphandle kohlelo lokuhlola, ukuqapha, izindleko kanye nokuphepha kuyindlela yokuthola izimanga ezimbi. Ukuqina okufanayo okusebenzisa kunoma iyiphi insizakalo ebalulekile ye-.NET kumele kufinyelele kungqimba yakho ye-ejenti, kuhambisane nesimo sokungenzeka sama-LLM.

Qala ngokuhlola amathuluzi nezindlela zokuhlanganisa ngeyunithi yakudala kanye nokuhlolwa kokuhlanganiswa ngaphambi kokukhathazeka ngokuziphatha kwemodeli. Wonke umsebenzi we-C# ongabizwa yi-ejenti kufanele uhlolwe ngokuzimela, ngokufakwayo okuqinisekile kanye nemiphumela. Bese uklama izikripthi zengxoxo ezilawulwayo ezisebenzisa izindlela ezigcwele zokuxhumana, ukuqinisekisa hhayi nje impendulo yokugcina kodwa futhi ukuthi yimaphi amathuluzi abizwa nokuthi isimo savela kanjani.

Ukubonwa kufanele kulandelele ukubambezeleka, ukusetshenziswa kwamathokheni kanye namazinga empumelelo kuzo zonke izindlela zokufeza ezihlukene. Kuwusizo kakhulu ukukala amathokheni okushesha kanye nokuqedwa ngokusebenzisana ngakunye, ahlukaniswe ngokuhamba komsebenzi, ithuluzi noma uhlobo lomsebenzisi, ukuze ukwazi ukubona ukuhlehla kanye nokwenyuka kwezindleko. Izingxoxo ezinde zibiza kakhulu, ngakho-ke tshala imali ekufinyezweni okuzenzakalelayo kanye namasu okunciphisa ahlakaniphile ukuze ugcine izimo zilula.

Ukuphepha akuxoxiswana ngakho uma ama-ejenti akho ethinta idatha ebucayi noma yamakhasimende. Kufanele usebenzise ukulawula okuqinile kokufinyelela kokuthi yimaphi amathuluzi namasethi wedatha i-ejenti engawabona, ubhalise yonke invoyisi yamathuluzi ngezinjongo zokuhlola futhi usebenzise zonke izingcingo zangaphandle ngezendlalelo zokuhlanza. Iziqinisekiso akufanele zifakwe kukhodi; zithembele kubunikazi obuphethwe, izitolo eziyimfihlo kanye nemikhuba evamile yokuphepha kwamafu osuvele uyisebenzisa kumasevisi amancane angewona e-AI.

Izimfuneko zokuthobela imithetho zithinta nendlela ogcina futhi ucubungula ngayo umlando wengxoxo. Ngoba amaseshini nezingxoxo zingase zibe nolwazi oluhlonza umuntu noma okuqukethwe okuyimfihlo, chaza izinqubomgomo zokugcina, amasu okungaziwa kanye nemithetho yokunciphisa idatha kusenesikhathi. Ikhono lokufaka amaseshini e-ejenti ku-series kanye nokususa i-seriali linamandla, kodwa kumele lilinganiswe nezibopho zomthetho nezokulawula.

Ngasohlangothini lwezindleko, ungawuthathi kancane umthelela wokungasebenzi kahle okuncane kakhulu. Izinguquko ezincane ngobukhulu obusheshayo, imvamisa yezingcingo zamathuluzi noma inani lama-ejenti ahambisanayo kungahumusha kube yizikweletu ezinkulu zanyanga zonke. Ukufaka amathuluzi ohlelweni, ukubuyekeza njalo izimpendulo ze-telemetry kanye nokulungisa, izinqubomgomo zememori kanye nokukhetha amamodeli kubalulekile ukugcina izindleko zizinzile ngokuhamba kwesikhathi.

Ukufakwa kanye nokwelulwa kulula uma uhlukanisa indiza yokulawula (lapho ulungiselela khona ama-ejenti kanye nemisebenzi yokusebenza) kusuka kundiza yokuphetha (lapho izingcingo zangempela zemodeli zisebenza khona). Ukuhlelwa okusekelwe esitsheni, imigqa yemiyalezo yemisebenzi eqhubekayo isikhathi eside kanye nezinsizakalo zamafu eziphethwe zokusingatha i-LLM konke kunegalelo ekuqineni. Imiphumela ingabe isingena kumadeshibhodi noma kumathuluzi e-BI njenge-Power BI ukuvala iluphu yempendulo yokuhlaziya nokukhombisa inani lebhizinisi.

Amathuluzi ahlanganisiwe njenge-AI Toolkit kanye nezandiso ze-Azure AI Foundry ze-Visual Studio Code zingenza kube lula kakhulu kulo mjikelezo wokuphila. Ngaphakathi komhleli ungahlola amakhathalogi amamodeli, usebenzise amamodeli aphethwe yi-GitHub noma endawo nge-Ollama, uqhathanise imiphumela eceleni, wakhe futhi usebenzise abahloli, ubone ngeso lengqondo imiphumela ku-Data Wrangler, ama-ejenti okuklama anezikhuthazo zesistimu, unamathisele amaseva e-MCP ukuze kuhlanganiswe amathuluzi kanye nokusebenzisana kwama-ejenti okulungisa amaphutha. I-Azure AI Foundry yengeza abaklami ababonakalayo, ukuvumelanisa kwe-YAML, ukukhiqizwa kwekhodi yokufinyelela kumamodeli e-Azure kanye nokuhlanganiswa kweklasi yokuqala kwamathuluzi afana ne-Bing Search kanye nabahumushi bekhodi.

Uma uhlanganisa lezi zithako—ukwakheka kwe-ejenti okuqinile, ukuphathwa kwesimo okucatshangelwe kahle, amathuluzi aqinile, imisebenzi yokusebenza esekelwe kugrafu lapho kudingeka khona, ukubonwa okujulile kanye nokusetshenziswa kwamafu—uthola ama-ejenti e-C# AI angewona nje ama-demo ahlakaniphile kodwa futhi ayizingxenye ezithembekile zezinhlelo zebhizinisi elikhulu. Ngokuklama ngokucophelela kanye nokusetshenziswa okufanele kwe-Azure OpenAI Assistants kanye ne-Microsoft Agent Framework, lawo ma-ejenti angathuthukisa ukusebenza kahle, ikhwalithi yolwazi kanye nokwenza izinto ngokuzenzakalela enhlanganweni yakho ngenkathi ehlala enakekelwa futhi ephephile.