Ukubonisana Okuzenzakalelayo kuhlola ukusetshenziswa kwereferensi yokubhala kabusha ye-chatbot

Namuhla, sishicilela isampula ye-chatbot entsha yomthombo ovulekile ebonisa indlela yokusebenzisa impendulo evela kukuhlola Okuzenzakalelayo Kokubonisana ukuze uphindaphinde okuqukethwe okukhiqizwa, ubuze imibuzo ecacisayo, futhi ufakazele ukulunga kwempendulo.

Ukusetshenziswa kwe-chatbot kuphinde kukhiqize ilogi yokuhlola ehlanganisa izincazelo eziqinisekiswa ngokwezibalo zokufaneleka kwempendulo kanye nesixhumi esibonakalayo somsebenzisi esibonisa onjiniyela inqubo ephindaphindwayo, yokuphinda ibhale okwenzeka ngemuva kwezigcawu. Ukuhlola Okuzenzakalelayo Kokubonisana kusebenzisa ukukhipha okunengqondo ukuze kubonise ngokuzenzakalelayo ukuthi isitatimende silungile. Ngokungafani namamodeli amakhulu olimi, amathuluzi Okubonisana Okuzenzakalelayo awaqageli noma abikezela ukunemba. Kunalokho, bathembele ebufakazini bezibalo ukuze baqinisekise ukuthotshelwa kwezinqubomgomo. Lokhu okuthunyelwe kwebhulogi kungena ngokujulile ekusetshenzisweni kwezakhiwo ze-Automated Reasoning ihlola ukubhala kabusha i-chatbot.

Thuthukisa ukunemba nokungafihli lutho ngamasheke Okuzenzakalelayo Okubonisana

Ngezinye izikhathi ama-LLM angakwazi ukukhiqiza izimpendulo ezizwakala zikholisa kodwa eziqukethe amaphutha ayiqiniso—into eyaziwa ngokuthi ukubona izinto ezingekho. Ukuhlola Okuzenzakalelayo Kokubonisana kuqinisekisa umbuzo womsebenzisi kanye nempendulo ekhiqizwe yi-LLM, kunikeza impendulo ebhala kabusha ekhomba izitatimende ezingacacile, ukugomela okubanzi kakhulu, kanye nezimangalo ezingelona iqiniso ezisekelwe olwazini lweqiniso eliyisisekelo elifakwe ikhodi kuzinqubomgomo Zokubonisana Okuzenzakalelayo.

I-chatbot esebenzisa ukuhlola Okuzenzakalelayo Kokubonisana ukuze iphindaphinde izimpendulo zayo ngaphambi kokuzethula kubasebenzisi iyasiza ngcono ukunemba ngoba ingenza izitatimende ezinembayo eziphendula imibuzo kayebo/cha yabasebenzisi ngaphandle kokushiya indawo yokungacaci; futhi isiza ukuthuthukisa ukubonakala ngoba inganikeza ubufakazi obungaqinisekiswa ngokwezibalo bokuthi kungani izitatimende zayo zilungile, yenze izinhlelo zokusebenza ze-AI ezikhiqizayo zihloleke futhi zichazeke ngisho nasezindaweni ezilawulwayo.

Manje njengoba usuqonda izinzuzo, ake sihlole ukuthi ungakusebenzisa kanjani lokhu ezinhlelweni zakho zokusebenza.

Ukuqaliswa kwereferensi ye-Chatbot

I-chatbot iyi-Flask application edalula ama-API ukuze athumele imibuzo futhi ahlole isimo sempendulo. Ukuze ubonise ukusebenza kwangaphakathi kwesistimu, ama-API aphinde avumele ukuthi ubuyise ulwazi mayelana nesimo sokuphindaphinda ngakunye, impendulo evela ekuhloleni Okuzenzakalelayo Kokubonisana, kanye nokwaziswa kokubhala kabusha okuthunyelwa ku-LLM.

Ungasebenzisa i-frontend ye-NodeJS yohlelo lokusebenza ukuze ulungiselele i-LLM evela ku-Amazon Bedrock ukuze ukhiqize izimpendulo, ukhethe inqubomgomo Yokubonisana Okuzenzakalelayo ukuze uqinisekiswe, futhi usethe inani eliphezulu lokuphindaphinda ukuze ulungise impendulo. Ukukhetha uchungechunge lwengxoxo kusixhumi esibonakalayo somsebenzisi kuvula iphaneli yokususa iphutha kwesokudla ebonisa impinda ngayinye kokuqukethwe kanye nokuphumayo kokuqinisekisa.

Uma Ukubonisana Okuzenzakalelayo kuhlola ukuthi impendulo ivumelekile, incazelo yokufaneleka iyaboniswa.

Umfanekiso 2 – Ukubonisana Okuzenzakalelayo kuhlola ubufakazi bokufaneleka

Isebenza kanjani iluphu yokubhala kabusha ephindaphindwayo

Ukusetshenziswa kwereferensi yomthombo ovulekile kusiza ngokuzenzakalelayo ukuthuthukisa izimpendulo ze-chatbot ngokuphindaphinda impendulo evela ekuhloleni Ukucabanga Okuzenzakalelayo nokubhala kabusha impendulo. Lapho ucelwa ukuthi uqinisekise umbuzo we-chatbot kanye nempendulo (i-Q&A), ukuhlola Okuzenzakalelayo Kokubonisana kubuyisela uhlu lokutholiwe. Ukuthola ngakunye kumele isitatimende esinengqondo esizimele esikhonjwe ku-Q&A yokufaka. Isibonelo, ku-Q&A “Sibiza malini isitoreji se-S3? E-US East (N. Virginia), i-S3 ibiza u-$0.023/GB nge-50Tb yokuqala; e-Asia Pacific (Sydney), i-S3 ibiza u-$0.025/GB ngo-50Tb wokuqala” Ukuhlola Okuzenzakalelayo Okubonisana kuzoveza okutholiwe okubili, okukodwa okuqinisekisa intengo ye-S2-03 kuthi- $01 kuthi i-$ 01 kithi. ap-eningizimu-mpumalanga-2.

Lapho kucutshungulwa okutholiwe kwe-Q&A, ukuhlola Okuzenzakalelayo Kokubonisana kwehlukanisa okokufaka kuhlu lwezakhiwo eziyiqiniso kanye nezimangalo ezenziwe ngokumelene nalezo zakhiwo. Isisekelo singaba isitatimende esiyiqiniso embuzweni womsebenzisi, njengokuthi “Ngingumsebenzisi we-S3 eVirginia,” noma umcabango obekwe empendulweni, njengokuthi “Okwezicelo ezithunyelwe kithi-empumalanga-1…” Isimangalo simelela isitatimende esiqinisekiswayo. Esibonelweni sethu sentengo ye-S3 kusukela esigabeni sangaphambilini, iSifunda singaba isisekelo, futhi iphoyinti lentengo lizoba isimangalo.

Ukuthola ngakunye kuhlanganisa umphumela wokuqinisekisa (VALID, INVALID, SATISFIABLE, TRANSLATION_AMBIGUOUS, IMPOSSIBLE) kanye nempendulo edingekayo ukuze ubhale kabusha impendulo ukuze ibe njalo VALID. Impendulo iyashintsha kuye ngomphumela wokuqinisekisa. Isibonelo, okutholiwe okungacacile kuhlanganisa ukutolika okubili kombhalo ofakiwe, okutholiwe okwanelisayo kufaka phakathi izimo ezimbili ezibonisa ukuthi izimangalo zingaba yiqiniso kanjani kwezinye izimo futhi zingamanga kwezinye. Ungabona izinhlobo ezingaba khona zokuthola emibhalweni yethu ye-API.

Njengoba lo mongo ungasekho endleleni, singangena sijule ekutheni ukuqaliswa kwesithenjwa kusebenza kanjani:

Impendulo yokuqala kanye nokuqinisekisa

Lapho umsebenzisi ehambisa umbuzo nge-UI, uhlelo lokusebenza luqale lushayele i-Bedrock LLM emisiwe ukuze lukhiqize impendulo, bese lushayela i-ApplyGuardrail API ukuze iqinisekise i-Q&A.

Ukusebenzisa okukhiphayo okuvela ku-Automated Reasoning hlola ku- ApplyGuardrail impendulo, uhlelo lokusebenza lungena ku-loop lapho ukuphindaphinda ngakunye kuhlola Ukucabanga Okuzenzakalelayo kuhlola impendulo, lenze isenzo esifana nokucela i-LLM ukuthi ibhale kabusha impendulo ngokusekelwe kumpendulo, bese ishaya ucingo. ApplyGuardrail ukuze uqinisekise okuqukethwe okubuyekeziwe futhi.

Iluphu yokubhala kabusha (Inhliziyo yesistimu)

Ngemva kokuqinisekisa kokuqala, isistimu isebenzisa okukhiphayo okuvela ekuhloleni Okuzenzakalelayo Kokubonisana ukuze inqume isinyathelo esilandelayo. Okokuqala, ihlunga okutholakele ngokusekelwe kokubalulekile kwakho – ibhekana nokubaluleke kakhulu kuqala: TRANSLATION_AMBIGUOUS, IMPOSSIBLE, INVALID, SATISFIABLE, VALID. Bese, ikhetha ukuthola okubalulekile okuphezulu futhi ibhekane nakho ngomqondo ongezansi. Kusukela VALID igcina ohlwini olubekwe phambili, isistimu izokwamukela okuthile njenge VALID ngemuva kokubhekana neminye imiphumela.

• Ngoba TRANSLATION_AMBIGUOUS okutholakele, ukuhlola Okuzenzakalelayo Kokubonisana kubuyisela izincazelo ezimbili zombhalo ofakiwe. Ngoba SATISFIABLE Okutholakele, ukuhlola Okuzenzakalelayo Kokubonisana kubuyisela izimo ezimbili ezifakazela futhi ziphikise izimangalo. Isebenzisa impendulo, uhlelo lokusebenza lucela i-LLM ukuthi inqume ukuthi ifuna ukuzama futhi ibhale kabusha impendulo ukuze icacise okungaqondakali noma ibuze umsebenzisi imibuzo yokulandelela ukuze iqoqe ulwazi olwengeziwe. Ngokwesibonelo, i SATISFIABLE impendulo ingase ithi intengo ka-$0.023 ivumeleke kuphela uma iSifunda siyi-US East (N. Virginia). I-LLM ingasebenzisa lolu lwazi ukubuza mayelana neSifunda sesicelo. Lapho i-LLM inquma ukubuza imibuzo yokulandelela, iluphu iyama futhi ilinde umsebenzisi ukuthi aphendule imibuzo, bese i-LLM ikhiqiza kabusha impendulo ngokusekelwe ekucaciseni bese iluphu iqala kabusha.
• Ngoba IMPOSSIBLE okutholakele, i-Automated Reasoning ihlola ibuyisela uhlu lwemithetho ephikisana nezakhiwo – amaqiniso amukelekile kokuqukethwe okokufaka. Ngokusebenzisa impendulo, uhlelo lokusebenza lucela i-LLM ukuthi ibhale kabusha impendulo ukuze igweme ukungqubuzana okunengqondo.
• Ngoba INVALID okutholakele, ukuhlola Okuzenzakalelayo Kokubonisana kubuyisela imithetho evela kunqubomgomo Yokubonisana Okuzenzakalelayo eyenza izimangalo zingavumelekile ngokusekelwe ezakhiweni nemithetho yenqubomgomo. Ngokusebenzisa impendulo, uhlelo lokusebenza lucela i-LLM ukuthi ibhale kabusha impendulo yayo ukuze ihambisane nemithetho.
• Ngoba VALID okutholakele, uhlelo lokusebenza luphuma ku-loop bese lubuyisela impendulo kumsebenzisi.

Ngemuva kokubhalwa kabusha kwempendulo ngayinye, uhlelo luthumela i-Q&A ku- ApplyGuardrail I-API yokuqinisekisa; ukuphindaphinda okulandelayo kweluphu kuqala ngempendulo esuka kule kholi. Ukuphindaphinda ngakunye kugcina okutholiwe kanye nokwaziswa okunomongo ogcwele kusakhiwo sedatha yochungechunge, kudala umkhondo wokuhlola wokuthi uhlelo lufike kanjani empendulweni eqondile.

Ukuqalisa nge-Automated Reasoning ihlola ukubhala kabusha i-chatbot

Ukuze uzame ukusebenzisa ireferensi yethu, isinyathelo sokuqala siwukudala inqubomgomo yokubonisana okuzenzakalelayo:

1. Zulazulela ku I-Amazon Bedrock ku-AWS Management Console kwesinye sezifunda ezisekelwayo e-United States noma ezifundeni zase-Europe.
2. Kusukela ekuzulazuleni kwesokunxele, vula i- Ukubonisana Okuzenzakalelayo ikhasi ku Yakha isigaba.
3. Ukusebenzisa imenyu eyehlayo ye- Dala inqubomgomo inkinobho, khetha Dala inqubomgomo yesampula.
4. Faka igama lenqubomgomo bese ukhetha Dala inqubomgomo ngaphansi kwekhasi.

Uma usudale inqubomgomo, ungaqhubeka ukuze ulande futhi usebenzise ukusetshenziswa kwesithenjwa:

5. Clone inqolobane ye-Amazon Bedrock Samples.
6. Landela imiyalelo kufayela le-README ukuze ufake okuncikile, wakhe indawo engaphambili, bese uqala uhlelo lokusebenza.
7. Usebenzisa isiphequluli osithandayo zulazulela uye futhi uqale ukuhlola.

Gcina imininingwane yokusetshenziswa

Uma uhlela ukulungisa lokhu kuqaliswa ukuze kusetshenziswe ukukhiqiza, lesi sigaba sidlula izingxenye ezibalulekile ekwakhiweni kwe-backend. Uzothola lezi zingxenye kumkhombandlela ongemuva wendawo yokugcina.

• I-ThreadManager: Ihlela ukuphathwa komjikelezo wempilo yengxoxo. Iphatha ukudalwa, ukubuyisa, kanye nokulandelelwa kwesimo kochungechunge lwezingxoxo, igcina isimo esifanele kuyo yonke inqubo yokubhala kabusha. I-ThreadManager isebenzisa imisebenzi evikelekile ngochungechunge isebenzisa isikhiya ukuze isize ukuvimbela izimo zomjaho lapho imisebenzi eminingi izama ukulungisa ingxoxo efanayo ngesikhathi esisodwa. Iphinde ilandelele uchungechunge olulindele okokufaka komsebenzisi futhi ingahlonza imicu yakudala eyeqe isikhathi sokuvala esilungisekayo.
• I-ThreadProcessor: Iphatha iluphu yokuphinda ibhale kusetshenziswa iphethini yomshini wombuso ukuze kugeleze okucacile, nokugcinwa kokulawula. Iphrosesa ilawula ukushintshwa kombuso phakathi kwezigaba ezifana GENERATE_INITIAL, VALIDATE, CHECK_QUESTIONS, HANDLE_RESULTfuthi REWRITING_LOOPeqhubekisela phambili ingxoxo ngendlela efanele esigabeni ngasinye.
• Isevisi yokuqinisekisa: Iihlanganiswe ne-Amazon Bedrock Guardrails. Le sevisi ithatha impendulo ngayinye ekhiqizwe yi-LLM futhi iyithumele ukuze iqinisekiswe kusetshenziswa i- ApplyGuardrail I-API. Iphatha ukuxhumana ne-AWS, ilawula ukucabanga kokuzama kabusha ngokubuyisela emuva kokungaphumeleli kwesikhashana, futhi ihlukanise imiphumela yokuqinisekisa ibe yimiphumela ehleliwe.
• LLMResponseParser: Ichaza izinhloso ze-LLM ngesikhathi sokubhala kabusha iluphu. Uma isistimu icela i-LLM ukuthi ilungise impendulo engavumelekile, imodeli kufanele inqume ukuthi izozama yini ukubhala kabusha (REWRITE), buza imibuzo ecacisayo (ASK_QUESTIONS), noma amemezele ukuthi umsebenzi awunakwenzeka ngenxa yezakhiwo eziphikisanayo (IMPOSSIBLE). Umhlaziyi uhlola impendulo ye-LLM kumaka athile afana nokuthi “DECISION:“,”ANSWER:“, futhi”QUESTION:“, ikhipha ulwazi oluhlelekile kokuphumayo kolimi lwemvelo. Isingatha ukufometha kokumaka phansi kahle futhi ibeka imikhawulo enanini lemibuzo (ubuningi obungu-5).
• I-AuditLogger: Ibhala amalogi ahlelekile e-JSON kufayela lokungena lokuhlola elizinikele, irekhoda izinhlobo ezimbili ezibalulekile zemicimbi: VALID_RESPONSE lapho impendulo idlula ukuqinisekiswa, futhi MAX_ITERATIONS_REACHED uma isistimu iqeda inombolo emisiwe yemizamo yokuzama futhi. Ukufakwa ngakunye kocwaningo kuthwebula isitembu sesikhathi, i-ID yochungechunge, ukwaziswa, impendulo, i-ID yemodeli, nokutholwe kokuqinisekisa. Umgawuli uphinda akhiphe futhi arekhode ukushintshisana kwe-Q&A kusukela ekucaciseni okuphindwaphindwayo, okuhlanganisa ukuthi umsebenzisi uphendule noma weqile yini imibuzo.

Ndawonye, ​​lezi zingxenye zisiza ukudala isisekelo esiqinile sokwakha izinhlelo zokusebenza ezithembekile ze-AI ezihlanganisa ukuguquguquka kwamamodeli amakhulu olimi nokuqina kokuqinisekisa kwezibalo.

Ukuze uthole umhlahlandlela onemininingwane wokusebenzisa ukuhlola Okuzenzakalelayo Kokubonisana ekukhiqizeni:

Mayelana nababhali