Impendulo emfushane: Ukuze wakhe i-ejenti ye-AI esebenza ngokusebenza, yiphathe njenge-loop elawulwayo: thatha okokufaka, unqume isenzo esilandelayo, shayela ithuluzi elinobubanzi obuncane, bheka umphumela, bese uphinda kuze kudlule ukuhlolwa okucacile "okuqediwe". Ithola ukugcinwa kwayo lapho umsebenzi unezinyathelo eziningi futhi uqhutshwa yithuluzi; uma isiphakamiso esisodwa siwuxazulula, yeqa i-ejenti. Engeza ama-schema amathuluzi aqinile, imikhawulo yesinyathelo, ukuloga, kanye nomqinisekisi/umgxeki ukuze lapho amathuluzi ehluleka noma okufakwayo kungacaci, i-ejenti ikhuphuke esikhundleni sokuyilupha.
Izinto ezibalulekile okufanele uzicabangele:
I-loop yokulawula: Sebenzisa okokufaka→ ukwenza→ qaphela ukuphindaphinda ngezimo zokuma ezicacile kanye nezinyathelo eziphezulu.
Ukuklama amathuluzi: Gcina amathuluzi emincane, ethayishiwe, evunyelwe, futhi eqinisekisiwe ukuze kuvinjelwe isiphithiphithi "sokwenza noma yini".
Ukuhlanzeka kwenkumbulo: Sebenzisa isimo sesikhathi esifushane kanye nokubuyisa isikhathi eside; gwema ukulahla imibhalo ephelele.
Ukumelana nokusebenzisa kabi: Engeza uhlu lwabavunyelwe, imikhawulo yamanani, ukunganaki, kanye "nokungaqali" kwezenzo eziyingozi.
Ukuhlolwa: Gcina isethi yesimo (ukwehluleka, ukungacaci kahle, imijovo) bese uphinda uyisebenzise njalo uma kushintsha.
🔗 Ungakala kanjani ukusebenza kwe-AI
Funda izindlela zokulinganisa ezisebenzayo zokulinganisa isivinini, ukunemba, kanye nokuthembeka.
🔗 Ungakhuluma kanjani ne-AI
Sebenzisa izikhuthazo, umongo, kanye nokulandelela ukuze uthole izimpendulo ezingcono.
🔗 Indlela yokuhlola amamodeli e-AI
Qhathanisa amamodeli usebenzisa izivivinyo, amarubrikhi, kanye nemiphumela yomsebenzi wangempela.
🔗 Indlela yokuthuthukisa amamodeli e-AI
Thuthukisa ikhwalithi kanye nezindleko ngokulungisa, ukucheba, kanye nokuqapha.
1) Ukuthi i-ejenti ye-AI iyini, ngokomuntu ojwayelekile 🧠
I-ejenti ye-AI iyi-loop. Amadokhumenti e-LangChain “Agents”
Yilokho kuphela. Iluphu enobuchopho phakathi.
Okufakiwe → cabanga → yenza → qaphela → phinda. Phinda wenze iphepha (isizathu + isenzo)
Kuphi:
-
Okokufaka kuyisicelo somsebenzisi noma umcimbi (i-imeyili entsha, ithikithi lokusekela, i-sensor ping).
-
Ukucabanga kuyisibonelo solimi sokucabanga ngesinyathelo esilandelayo.
-
Umthetho ubiza ithuluzi (sesha amadokhumenti angaphakathi, sebenzisa ikhodi, dala ithikithi, bhala impendulo). Umhlahlandlela wokushaya we-OpenAI Function
-
I-Observe ifunda umphumela wethuluzi.
-
Ukuphindaphinda yingxenye eyenza kuzwakale “kunobungcweti” esikhundleni sokuthi “kukhuluma kakhulu”. Amadokhumenti e-LangChain “Agents”
Amanye ama-ejenti empeleni angama-macro ahlakaniphile. Amanye asebenza njengomqhubi omncane ongakwazi ukuhlanganisa imisebenzi futhi alulame emaphutheni. Womabili ayabalwa.
Futhi, awudingi ukuzimela okugcwele. Eqinisweni... cishe awukufuni 🙃
2) Kunini lapho kufanele wakhe khona i-ejenti (futhi nini lapho kungafanele wakhe khona) 🚦
Yakha i-ejenti uma:
-
Umsebenzi unezinyathelo eziningi futhi uyashintsha kuye ngokuthi kwenzekani phakathi nendawo.
-
Umsebenzi udinga ukusetshenziswa kwamathuluzi (izizindalwazi, ama-CRM, ukwenziwa kwekhodi, ukukhiqizwa kwamafayela, iziphequluli, ama-API angaphakathi). Amadokhumenti e-LangChain “Amathuluzi”
-
Ufuna imiphumela ephindaphindwayo enezivikelo, hhayi izimpendulo ezifika kanye kuphela.
-
Ungachaza ukuthi “kwenziwe” ngendlela ikhompyutha engayihlola ngayo, ngisho noma ingacacile.
Ungakhi i-ejenti uma:
-
Impendulo elula esheshayo + impendulo iyakuxazulula (ungazenzisi ngokweqile, uzozizonda kamuva).
-
Udinga ukuzimisela okuphelele (ama-ejenti angaba njalo, kodwa hhayi amarobhothi).
-
Awunawo amathuluzi noma idatha yokuxhuma - ngakho-ke ikakhulukazi kumane kuyimizwa nje.
Masikhulume iqiniso: ingxenye "yamaphrojekthi e-AI agent" ingaba umsebenzi onemithetho embalwa yokuhlanganisa. Kodwa-ke, ngezinye izikhathi i-vibe nayo ibalulekile 🤷♂️
3) Yini eyenza inguqulo enhle ye-AI agent ✅
Nasi isigaba esithi “Yini eyenza inguqulo enhle” osicelile, ngaphandle kokuthi ngizokhuluma ngokungagwegwesi:
Inguqulo enhle ye-AI agent akuyona leyo ecabanga kakhulu. Yileyo:
-
Uyazi ukuthi yini evunyelwe ukuyenza (imingcele yesikhala)
-
Isebenzisa amathuluzi ngokwethembeka (izingcingo ezihleliwe, ukuzama kabusha, ukuphelelwa yisikhathi) Umhlahlandlela wokushaya ucingo we-OpenAI Function AWS “Ukuphelelwa yisikhathi, ukuzama kabusha, kanye nokubuyela emuva nge-jitter”
-
Igcina isimo sihlanzekile (inkumbulo engaboli) I-LangChain “Ukubuka konke kwememori”
-
Ichaza izenzo zayo (izindlela zokuhlola, hhayi izithiyo eziyimfihlo zokucabanga) I-NIST AI RMF 1.0 (ukuthembeka kanye nokucaca)
-
Ima ngendlela efanele (ukuhlolwa kokuqedwa, izinyathelo eziphezulu, ukukhushulwa) amadokhumenti e-LangChain “Agents”
-
Uhluleka ngokuphepha (ucela usizo, akalimazi igunya) I-NIST AI RMF 1.0
-
Iyahlolwa (ungayiqhuba ezimweni ezikheniwe bese uthola imiphumela)
Uma i-ejenti yakho ingakwazi ukuhlolwa, empeleni iyi-slot machine eqiniseka kakhulu. Kumnandi emaphathini, kuyasabisa ekukhiqizeni 😬
4) Izisekelo zokwakha ze-ejenti ("i-anatomy" 🧩)
Ama-agent amaningi aqinile analezi zingcezu:
A) Iluphu yokulawula 🔁
Lona ngumhleli wezinhlelo:
-
thatha umgomo
-
cela imodeli yesenzo esilandelayo
-
ithuluzi lokusebenzisa
-
faka ukubonwa
-
phinda kuze kube yilapho kuqediwe amadokhumenti e-“Agents” e-LangChain
B) Amathuluzi (okwaziwa nangokuthi amakhono) 🧰
Amathuluzi yiwona enza i-ejenti isebenze kahle: amadokhumenti e-“Tools” e-LangChain
-
imibuzo yesizindalwazi
-
ukuthumela ama-imeyili
-
ukudonsa amafayela
-
ikhodi esebenzayo
-
ukubiza ama-API angaphakathi
-
ukubhalela amaspredishithi noma ama-CRM
C) Inkumbulo 🗃️
Izinhlobo ezimbili zibalulekile:
-
inkumbulo yesikhathi esifushane: umongo wokusebenza kwamanje, izinyathelo zakamuva, uhlelo lwamanje
-
inkumbulo yesikhathi eside: okuthandwa ngumsebenzisi, umongo wephrojekthi, ulwazi olutholiwe (ngokuvamile ngokushumeka + isitolo sevektha) Iphepha le-RAG
D) Inqubomgomo yokuhlela kanye nezinqumo 🧭
Ngisho noma ungakubizi ngokuthi “ukuhlela”, udinga indlela:
-
uhlu lokuhlola
-
Iphepha le-ReAct elithi “cabanga bese kuba ithuluzi” lesitayela se-ReAct
-
amagrafu emisebenzi
-
amaphethini abaphathi nabasebenzi
-
amaphethini omphathi-isisebenzi i-Microsoft AutoGen (uhlaka lwama-ejenti amaningi)
E) Izithiyo zokuvikela nokuhlola 🧯
-
izimvume
-
izikimu zamathuluzi aphephile Imiphumela Ehlelekile ye-OpenAI
-
ukuqinisekiswa kokukhipha
-
imikhawulo yesinyathelo
-
ukuqopha
-
ihlola i- NIST AI RMF 1.0
Yebo, kungubunjiniyela ngaphezu kokukhuthaza. Okuyi... iphuzu elibalulekile.
5) Ithebula Lokuqhathanisa: izindlela ezidumile zokwakha i-ejenti 🧾
Ngezansi kukhona "Ithebula Lokuqhathanisa" elingokoqobo - elinezimo ezimbalwa ezingavamile, ngoba amaqembu angempela ayizinto ezingavamile 😄
| Ithuluzi / Uhlaka | Izithameli | Intengo | Kungani kusebenza | Amanothi (isiphithiphithi esincane) | |
|---|---|---|---|---|---|
| I-LangChain | abakhi abathanda izingxenye zesitayela se-lego | i-free-ish + i-infra | uhlelo olukhulu lwezinhlelo zokusebenza zamathuluzi, inkumbulo, amaketanga | ungashesha ukudla i-spaghetti uma ungasho izinto ngokucacile | |
| I-LlamaIndex | Amaqembu anzima kakhulu | i-free-ish + i-infra | amaphethini okubuyisa aqinile, ukukhomba, izixhumi | kuhle kakhulu uma i-ejenti yakho “ingukusesha + isenzo”… okuyinto evamile | |
| Indlela yesitayela sabasizi be-OpenAI | amaqembu afuna ukusethwa okusheshayo | okusekelwe ekusetshenzisweni | amaphethini okushaya amathuluzi akhelwe ngaphakathi kanye nesimo sokusebenza | akulula ukuzivumelanisa nezimo kwamanye amakhona, kodwa kuhlanzekile ezinhlelweni zokusebenza eziningi | I-OpenAI isebenzisa i-API ye-OpenAI Assistants function calling |
| I-Semantic Kernel | onjiniyela abafuna ukuhlelwa okuhlelekile | i-free-ish | isifinyezo esihle samakhono/imisebenzi | kuzwakala “kuhlelekile ebhizinisini” - ngezinye izikhathi lokho kuyancomeka 😉 | |
| I-AutoGen | abahloli bezinto eziningi | i-free-ish | amaphethini okubambisana phakathi kwe-ejenti ne-ejenti | angakhuluma ngokweqile; beka imithetho eqinile yokuqeda | |
| I-CrewAI | abalandeli “bamaqembu e-ejenti” | i-free-ish | izindima + imisebenzi + ukudluliselana kulula ukukuveza | isebenza kahle kakhulu uma imisebenzi ilula, hhayi ithambile | |
| I-haystack | sesha + abantu abasebenzisa amapayipi | i-free-ish | amapayipi aqinile, ukubuyisa, izingxenye | kungabi “i-ejenti yeshashalazi” kakhulu, kube “ifektri ewusizo” kakhulu | |
| Gingqa eyakho (iluphu eyenziwe ngokwezifiso) | ama-freak okulawula (onothando) | isikhathi sakho | umlingo omncane, ukucaca okuphezulu | ngokuvamile kuyindlela engcono kakhulu yesikhathi eside… uze uvuselele konke 😅 |
Akekho owinile oyedwa. Ukukhetha okungcono kuncike ekutheni umsebenzi oyinhloko we-ejenti yakho ukuthola kabusha, ukusebenzisa amathuluzi, ukuxhumanisa ama-ejenti amaningi, noma ukuzenzekela komsebenzi.
6) Indlela Yokwakha I-AI Agent isinyathelo ngesinyathelo (iresiphi yangempela) 🍳🤖
Lena yingxenye abantu abaningi abayigwemayo, bese bezibuza ukuthi kungani i-ejenti iziphethe njenge-raccoon epantry.
Isinyathelo 1: Chaza umsebenzi ngomusho owodwa 🎯
Izibonelo:
-
"Bhala impendulo yekhasimende usebenzisa inqubomgomo kanye nomongo wethikithi, bese ucela imvume."
-
"Phenya umbiko wesiphazamisi, uwukopishe, bese uphakamisa isixazululo."
-
"Guqula amanothi emihlangano angaphelele abe yimisebenzi, abanikazi, kanye nezinsuku zokugcina."
Uma ungakwazi ukukuchaza kalula, i-ejenti yakho nayo ayikwazi. Ngisho ukuthi ingakwenza, kodwa izokusungula, futhi ukwenza ngcono yilapho isabelomali siya khona.
Isinyathelo 2: Nquma izinga lokuzimela (eliphansi, eliphakathi nendawo, elinongwe) 🌶️
-
Ukuzimela okuphansi: kusikisela izinyathelo, ukuchofoza komuntu "kuyavunywa"
-
Okuphakathi: kusebenzisa amathuluzi, kukhipha okubhaliwe, kwandisa ukungaqiniseki
-
Okuphezulu: kusebenzisa kusukela ekuqaleni kuya ekugcineni, kucindezela abantu kuphela uma kuhlukile
Qala phansi kunalokho okufunayo. Ungahlala uvuselela kamuva.
Isinyathelo 3: Khetha isu lakho lemodeli 🧠
Ngokuvamile ukhetha:
-
imodeli eyodwa eqinile yayo yonke into (elula)
-
imodeli eyodwa eqinile + imodeli encane yezinyathelo ezishibhile (ukuhlela, ukuhambisa)
-
amamodeli akhethekile (umbono, ikhodi, inkulumo) uma kudingeka
Futhi nquma:
-
amathokheni aphezulu
-
izinga lokushisa
-
ukuthi uvumela yini imikhondo emide yokucabanga ngaphakathi (ungakwazi, kodwa ungavezi uchungechunge lwemicabango olungakalungi kubasebenzisi bokugcina)
Isinyathelo 4: Chaza amathuluzi ngama-schema aqinile 🔩
Amathuluzi kufanele abe:
-
okuncane
-
okuthayishiwe
-
kuvunyelwe
-
Imiphumela Ehleliwe ye-OpenAI eqinisekisiwe
Esikhundleni sethuluzi elibizwa ngokuthi do_anything(input: string), yenza:
-
search_kb(umbuzo: intambo) -> imiphumela[] -
dala_ithikithi(isihloko: intambo, umzimba: intambo, okubalulekile: i-enum) -> i-ticket_id -
send_email(kuya: intambo, isihloko: intambo, umzimba: intambo) -> isimoUmhlahlandlela wokushaya we-OpenAI Function
Uma unika i-ejenti i-chainsaw, ungathuki uma inciphisa uthango ngokususa nothango.
Isinyathelo 5: Yakha iluphu yokulawula 🔁
Iluphu encane:
-
Qala ngomgomo + umongo wokuqala
-
Buza imodeli: “Isenzo esilandelayo?”
-
Uma ucingo lwethuluzi - sebenzisa ithuluzi
-
Faka ukubonwa
-
Hlola isimo sokuma
-
Phinda (ngezinyathelo eziphezulu) amadokhumenti e-“Agents” e-LangChain
Engeza:
-
izikhathi zokuvala
-
ukuzama kabusha (qaphela - ukuzama kabusha kungalandela) I-AWS “Izikhathi zokuphelelwa yisikhathi, ukuzama kabusha, kanye nokubuyela emuva nge-jitter”
-
ukufometha iphutha lethuluzi (kucacile, kuhlelekile)
Isinyathelo 6: Engeza inkumbulo ngokucophelela 🗃️
Isikhathi esifushane: gcina "isifinyezo sesimo" esincane sibuyekeziwe isinyathelo ngasinye. "Ukubuka konke kwememori" kweLangChain
Isikhathi eside: gcina amaqiniso aqinile (izintandokazi zomsebenzisi, imithetho yenhlangano, amadokhumenti azinzile).
Dwebela ushiye isithupha:
-
uma kushintsha njalo - gcina isikhathi esifushane
-
uma kuzinzile - gcina isikhathi eside
-
uma kuzwela - gcina kancane (noma kungazweli nhlobo)
Isinyathelo 7: Engeza ukuqinisekiswa kanye nephasi "lomgxeki" 🧪
Iphethini eshibhile nesebenzayo:
-
i-ejenti ikhiqiza umphumela
-
isiqinisekisi sihlola isakhiwo kanye nemikhawulo
-
ukubuyekezwa kwemodeli yokugxeka okungakhethwa kwezinyathelo ezingekho noma ukwephulwa kwenqubomgomo i-NIST AI RMF 1.0
Akuphelele, kodwa kuthinta ubuningi obushaqisayo bokungasho lutho.
Isinyathelo 8: Bhala konke ozozisola ngakho ngokungabhalisi 📜
Ilogi:
-
izingcingo zamathuluzi + okokufaka + okukhiphayo
-
izinqumo ezenziwe
-
amaphutha
-
imiphumela yokugcina
-
amathokheni kanye nokubambezeleka kwe-OpenTemetry observability primer
Ikusasa - uzokubonga. Isipho - uzokhohlwa. Yilokho nje ukuphila 😵💫
7) Ukushaya amathuluzi okungakuphuli umphefumulo wakho 🧰😵
Ukubiza amathuluzi yilapho "Indlela Yokwakha I-AI Agent" iba khona ubunjiniyela besofthiwe bangempela.
Yenza amathuluzi athembeke (ukuthembeka kuhle)
Amathuluzi athembekile yilawa:
-
okunqunyiwe
-
ububanzi obuncane
-
kulula ukuyihlola
-
kuphephile ukuphinda usebenzise i-Stripe ethi “Idempotent requests”
Engeza izivikelo zokuvikela engqimbeni yamathuluzi, hhayi nje izixwayiso
Iziphakamiso ziyiziphakamiso ezinesizotha. Ukuqinisekiswa kwamathuluzi kuwumnyango okhiyiwe. Imiphumela Ehlelekile ye-OpenAI
Yenza:
-
uhlu lwabavunyelwe (amathuluzi angasebenza)
-
ukuqinisekiswa kokufaka
-
imikhawulo yesilinganiso I -OpenAI Umhlahlandlela wemikhawulo yesilinganiso
-
ukuhlolwa kwemvume ngomsebenzisi/inhlangano ngayinye
-
"Imodi yokugijima ngaphandle kokusebenzisa amandla" yezenzo eziyingozi
Umklamo wokwehluleka okuyingxenye
Amathuluzi ayahluleka. Amanethiwekhi ayazamazama. Ukuqinisekiswa kuyaphelelwa yisikhathi. I-ejenti kumele:
-
chaza amaphutha
-
zama futhi nge-backoff uma kufaneleka isu lokuzama kabusha le-Google Cloud (backoff + jitter)
-
khetha amathuluzi ahlukile
-
ukukhuphuka uma ubambekile
Icebo eliphumelelayo ngokuthula: buyisela amaphutha ahlelekile afana nalawa:
-
uhlobo: auth_error -
uhlobo: alutholakalanga -
uhlobo: rate_limited
Ngakho imodeli ingaphendula ngokuhlakanipha esikhundleni sokwethuka.
8) Inkumbulo ekusizayo esikhundleni sokukukhathaza 👻🗂️
Inkumbulo inamandla, kodwa ingaba futhi yi-junk drawer.
Inkumbulo yesikhathi esifushane: gcina incane
Sebenzisa:
-
izinyathelo zokugcina ze-N
-
isifinyezo esisebenzayo (sibuyekezwa njalo nge-loop)
-
uhlelo lwamanje
-
imikhawulo yamanje (isabelomali, isikhathi, izinqubomgomo)
Uma ulahla konke kumongo, uthola:
-
izindleko eziphakeme
-
ukubambezeleka okuhamba kancane
-
ukudideka okwengeziwe (yebo, noma kunjalo)
Inkumbulo yesikhathi eside: ukubuyisa phezu "kokugcwalisa"
Iningi "lenkumbulo yesikhathi eside" lifana kakhulu nalokhu:
-
ukushumeka
-
isitolo se-vector
-
iphepha le-RAG lokubuyisa isizukulwane esingeziwe (i-RAG)
I-ejenti ayigcini ngekhanda. Ithola izingcezu ezifanele kakhulu ngesikhathi sokusebenza. I-LlamaIndex “Isingeniso ku-RAG”
Imithetho yenkumbulo ewusizo
-
Gcina "izintandokazi" njengamaqiniso acacile: "Umsebenzisi uthanda izifinyezo zezinhlamvu futhi uyazonda ama-emoji" (lol, hhayi lapha kodwa 😄)
-
"Izinqumo" zesitolo ezinezitembu zesikhathi noma izinguqulo (ngaphandle kwalokho ukuphikisana kuyanqwabelana)
-
Ungalokothi ugcine izimfihlo ngaphandle kokuthi kufanele ngempela uzigcine
Nasi isifaniso sami esingaphelele: inkumbulo ifana nesiqandisi. Uma ungalokothi usihlanze, ekugcineni isangweji lakho linambitheka njengo-anyanisi nokuzisola.
9) Amaphethini okuhlela (kusukela kokulula kuya kokumangalisayo) 🧭✨
Ukuhlela kumane nje kuwukuwohloka okulawulwayo. Ungakwenzi kube yimfihlakalo.
Iphethini A: Umhleli wohlu lokuhlola ✅
-
Imodeli ikhipha uhlu lwezinyathelo
-
Isebenza isinyathelo ngesinyathelo
-
Isimo sohlu lokuhlola izibuyekezo
Kuhle kakhulu ekubhaliseni. Kulula, kuyahlolwa.
Iphethini B: Iluphu Yokwenza Kabusha (isizathu + isenzo) 🧠→🧰
-
imodeli inquma ucingo lwethuluzi elilandelayo
-
ibona umphumela
-
iphinda iphepha le-ReAct
Lona umuzwa we-ejenti wakudala.
Iphethini C: Umphathi-isisebenzi 👥
-
umphathi uhlukanisa umgomo ube yimisebenzi
-
abasebenzi benza imisebenzi ekhethekile
-
umphathi uhlanganisa imiphumela ye -Microsoft AutoGen (uhlaka lwama-ejenti amaningi)
Lokhu kuwusizo uma imisebenzi ingalinganiswa, noma uma ufuna "izindima" ezahlukene ezifana nalezi:
-
umcwaningi
-
umqondisi wekhodi
-
umhleli
-
Isihloli se-QA
Iphethini D: Hlela bese uqhuba ngokuhlela kabusha 🔄
-
dala uhlelo
-
qalisa
-
uma imiphumela yamathuluzi ishintsha iqiniso, hlela kabusha
Lokhu kuvimbela umenzi ukuthi alandele ngenkani icebo elibi. Abantu nabo bayakwenza lokhu, ngaphandle uma bekhathele, lapho-ke nabo balandela amacebo amabi.
10) Ukuphepha, ukuthembeka, kanye nokungaxoshwa 🔐😅
Uma i-ejenti yakho ingathatha izinyathelo, udinga ukwakheka kokuphepha. Akukuhle ukuba nakho. Isidingo. I-NIST AI RMF 1.0
Imikhawulo eqinile
-
izinyathelo eziphezulu ngokugijima ngakunye
-
izingcingo zamathuluzi eziphezulu ngomzuzu
-
imali eningi echithwe ngeseshini ngayinye (isabelomali sethokheni)
-
amathuluzi anqunyelwe ngemuva kokuvunyelwa
Ukuphathwa kwedatha
-
lungisa okokufaka okubucayi ngaphambi kokungena ngemvume
-
izindawo ezihlukene (ukuthuthukiswa vs ukukhiqizwa)
-
izimvume zethuluzi lelungelo elincane kakhulu
Imikhawulo yokuziphatha
-
phoqelela umenzeli ukuthi asho izingcezu zobufakazi bangaphakathi (hhayi izixhumanisi zangaphandle, izinkomba zangaphakathi kuphela)
-
dinga amafulegi okungaqiniseki lapho ukuzethemba kuphansi
-
kudinga "ukubuza umbuzo ocacisayo" uma okufakiwe kungacacile
I-ejenti ethembekile akuyona eqiniseka kakhulu. Yileyo eyaziyo ukuthi iqagela nini… bese isho kanjalo.
11) Ukuhlola nokuhlola (ingxenye wonke umuntu ayigwemayo) 🧪📏
Awukwazi ukuthuthukisa lokho ongakwazi ukukulinganisa. Yebo, lowo mugqa uyacasula, kodwa kuyiqiniso elicasulayo.
Yakha isethi yesimo
Dala amacala okuhlola angu-30-100:
-
izindlela ezijabulisayo
-
amacala onqenqema
-
Amacala “okwehluleka kwethuluzi”
-
izicelo ezingacacile
-
Imiyalelo yokuphikisana (imizamo yokujova ngokushesha) I-OWASP Eziyi-10 Eziphezulu zezinhlelo zokusebenza ze-LLM I-OWASP LLM01 Injection Prompt
Imiphumela yamaphuzu
Sebenzisa izilinganiso ezifana nalezi:
-
izinga lempumelelo yomsebenzi
-
isikhathi sokuqeda
-
izinga lokubuyisa iphutha lethuluzi
-
izinga lokubona izinto ezingekho (izimangalo ezingenabo ubufakazi)
-
izinga lokuvunyelwa komuntu (uma kumodi egadiwe)
Ukuhlolwa kokuhlehla kwemiyalelo namathuluzi
Noma nini lapho ushintsha:
-
uhlelo lwethuluzi
-
imiyalelo yesistimu
-
ingqondo yokuthola kabusha
-
ifomethi yememori
Sebenzisa i-suite futhi.
Ama-ejenti ayizilwane ezizwelayo. Njengezitshalo zasekhaya, kodwa abiza kakhulu.
12) Amaphethini okusabalalisa angancibilikisi isabelomali sakho 💸🔥
Qala ngesevisi eyodwa
-
i-API yesilawuli se-ejenti
-
izinsizakalo zamathuluzi ngemuva kwayo
-
ukuloga + ukuqapha OpenTelemetry observability primer
Engeza izilawuli zezindleko kusenesikhathi
-
imiphumela yokubuyisa i-caching
-
ukucindezela isimo sengxoxo ngezifinyezo
-
kusetshenziswa amamodeli amancane okuhambisa nokukhipha
-
ukukhawulela "imodi yokucabanga ejulile" ezinyathelweni ezinzima kakhulu
Ukukhetha okuvamile kwezakhiwo
-
isilawuli esingenasimo + isitolo sesimo sangaphandle (DB/redis)
-
izingcingo zamathuluzi azinangqondo lapho kungenzeka khona i-Stripe “Izicelo ezingabonakali”
-
umugqa wemisebenzi emide (ukuze ungagcini isicelo sewebhu sivuliwe unomphela)
Futhi: yakha "iswishi yokubulala". Ngeke uyidinge kuze kube yilapho uyidinga ngempela 😬
13) Amanothi okuvala - inguqulo emfushane mayelana nokuthi Ungayakha Kanjani I-Agent ye-AI 🎁🤖
Uma ungakhumbuli lutho olunye, khumbula lokhu:
-
Indlela Yokwakha I-Agent ye-AI imayelana nokwakha iluphu ephephile ezungeze imodeli. Amadokhumenti e-“Agents” e-LangChain
-
Qala ngomgomo ocacile, ukuzimela okuphansi, namathuluzi aqinile. Imiphumela Ehlelekile ye-OpenAI
-
Engeza inkumbulo ngokubuyisa, hhayi ukugcwaliswa komongo okungapheli. Iphepha le-RAG
-
Ukuhlela kungaba lula - uhlu lokuhlola kanye nokuhlela kabusha kuya kude.
-
Ukubhalisa nokuhlola kuguqula isiphithiphithi se-agent sibe yinto ongayithumela. I-OpenTelemetry observability primer
-
Ama-Guardrails afanele ikhodi, hhayi nje izixwayiso. I-OWASP Eziyi-10 Eziphezulu zezinhlelo zokusebenza ze-LLM
I-ejenti ayilona umlingo. Luhlelo oluvame ukwenza izinqumo ezinhle ngokwanele ukuba lube yigugu… futhi luvume ukuthi lunqotshiwe ngaphambi kokuba lubangele umonakalo. Lududuza buthule, ngandlela thile 😌
Futhi yebo, uma ukwakhela kahle, kuzwakala sengathi uqasha umfundi omncane wedijithali ongalali, ngezinye izikhathi othukayo, futhi othanda amaphepha. Ngakho-ke, empeleni ungumfundi oqeqeshwayo.
Isibonelo sangempela: Ukwakha i-ejenti ye-AI yokusekela yokulinganisa i-triage 🎫🤖
Isimo
Cabanga ngethimba elincane le-SaaS elithola amathikithi okusekela angu-120-180 ngesonto. Amathikithi amaningi awanzima, kodwa asathatha isikhathi: ukusetha kabusha amaphasiwedi, imibuzo yokukhokha, imibiko yamaphutha, izicelo zezici, kanye nemiyalezo ethi "ingabe lokhu kulindelekile?".
I-chatbot elula ingabhala izimpendulo, kodwa ayikwazi ukuhlola isimo se-akhawunti ngokuthembekile, ifune ulwazi, ihlukanise ukuphuthuma, noma inqume ukuthi umuntu udinga ukungenelela nini. Yilapho i-ejenti inengqondo khona.
Umgomo awukona ukufaka esikhundleni sokusekelwa ngokugcwele. Umgomo ukwakha i-ejenti engazimeli kakhulu efunda ithikithi elisha, iqoqe umongo, ibhale impendulo, bese ihambisa ithikithi iye emgqeni ofanele. Umuntu usavuma noma yini ebhekene namakhasimende.
Lokho okudingwa umsizi
Ukuze isebenze ngokuphepha, i-ejenti idinga isethi encane, elawulwayo yezinto ezifakwayo namathuluzi:
-
Umbhalo wethikithi ongenayo
-
Uhlobo lohlelo lwekhasimende, ubudala be-akhawunti, kanye nesimo sokukhokha sakamuva
-
I-log yokushintsha umkhiqizo yakamuva noma izehlakalo ezaziwayo
-
Izihloko zesikhungo sosizo sangaphakathi
-
Ithuluzi lokubuyekeza ithikithi elinamasimu anqunyelwe
-
Ithuluzi lokuphendula eliyisiqalo, hhayi ithuluzi lokuthumela i-imeyili
-
Inqubomgomo ecacile yokwenyusa izinga
Uhlu lwamathuluzi kufanele luhlale luncane ngamabomu:
-
isikhungo_sosizo_sokusesha(umbuzo)
-
isimo_sekhasimende(i-id_yekhasimende)
-
hlola_izehlakalo_ezaziwayo(indawo_yomkhiqizo)
-
isigaba_sethikithi_sokubuyekeza(id_yethikithi, isigaba, okubalulekile)
-
impendulo_esalungiswa(id_yethikithi, umbhalo_wempendulo)
-
khuphuka_ube_ngumuntu(i-id_yethikithi, isizathu)
Qaphela ukuthi yini engekho: ayikho ithuluzi elithi “ikhasimende lokubuyisela imali”, “ukuvala i-akhawunti”, noma “ukuthumela impendulo yokugcina”. Lezo zenzo ziyingozi kakhulu enguqulweni yokuqala.
Isibonelo semiyalelo
Ungummeli wokusekela womkhiqizo we-SaaS.
Umsebenzi wakho ukuhlukanisa amathikithi angenayo, ukuqoqa kuphela umongo odingekayo, ukubhala impendulo ephakanyisiwe, bese unquma ukuthi ithikithi kufanele lidluliselwe phambili yini.
Imithetho:
Ungathumeli izimpendulo ngqo kumakhasimende.
Sebenzisa isikhungo sosizo ngaphambi kokuphendula imibuzo yomkhiqizo.
Hlola isimo sekhasimende ngaphambi kokuphendula imibuzo yokukhokha, uhlelo, noma ukufinyelela.
Uma ikhasimende likhuluma ngezinsongo zomthetho, ukulahleka kwedatha, izinkinga zokuphepha, ukwehluleka kokukhokha, ukukhanselwa kwe-akhawunti, noma ulimi oluthukuthele, kuyashintsha kube ngumuntu.
Uma impendulo ingasekelwa okuqukethwe kwesikhungo sosizo okutholiwe noma idatha ye-akhawunti, sho ukuthi yini engekho bese uqhubeka.
Misa ngemva kwezingcingo zamathuluzi eziyi-6 esiphezulu.
Ithikithi "liqediwe" kuphela uma linesigaba, okubalulekile, isifinyezo sobufakazi, impendulo ebhaliwe, kanye nokuthi "kudingeka imvume yomuntu" noma "kukhushulwe".
Indlela yokuyihlola
Qala ngamathikithi okuhlola angu-30 ngaphambi kokuwaxhumanisa nabasebenzisi ababukhoma:
-
Amathikithi ayi-10 ajwayelekile, njengokusetha kabusha iphasiwedi, imikhawulo yohlelo, kanye nemibuzo eyisisekelo ethi “ngenzenjani?”
-
Amathikithi okukhokha angu-5
-
Imibiko yeziphazamiso ezi-5
-
Amathikithi ayi-5 angacacile anolwazi olungekho
-
Amathikithi ayi-5 ayingozi, njengezinkinga zokuphepha, izidingo zokubuyiselwa imali, kanye nezikhalazo ezithukuthele
Ngethikithi ngalinye, thola amaphuzu:
-
Ingabe ikhethe isigaba esifanele?
-
Ingabe isebenzise ithuluzi elifanele ngaphambi kokuphendula?
-
Ingabe igweme izimangalo ezingasekelwa?
-
Ingabe lokhu kwandise amathikithi ayingozi?
-
Ingabe uhlaka ludinga ukuhlelwa okunzima?
Ispredishithi elula yokudlula/yokwehluleka yanele ekuqaleni. Ungakhi ngokweqile uhlelo lokuhlola ngaphambi kokuthi wazi ukuthi i-ejenti iletha inani.
Umphumela
Umphumela obonisayo: Ngokusekelwe ekubekeni isikhathi amathikithi esampula angu-30 ngaphambi nangemva kokusebenzisa lo msebenzi, i-lead yokusekela ingalinganisa okulandelayo:
-
Isikhathi esimaphakathi sokulinganisa kokuqala sincishisiwe kusuka kumaminithi ayi-6 ngethikithi kuya kumasekhondi angama-90
-
Amathikithi angu-30 ahlelwe ngemizuzu engu-45 esikhundleni samahora angu-3
-
Amathikithi angu-27 kwangu-30 abekwe esigabeni esifanele
-
Amathikithi ayingozi angu-5 kwangu-5 anyuswe ngendlela efanele
-
Izimpendulo zamakhasimende eziyi-0 ezithunyelwe ngaphandle kwemvume yomuntu
Lezi zinombolo ziyisibonelo sokulinganisa, hhayi isilinganiso esiqinisekisiwe. Ukulinganisa kulula ukukuphinda: beka isikhathi seqoqo elifanayo lamathikithi okuhlola ngesandla, bese uwasebenzisa ku-ejenti bese uqhathanisa ukunemba kwesigaba, ukunemba kokukhushulwa, kanye nesikhathi sokuhlela.
Yini engase ihambe kabi
I-ejenti isengahluleka ngezindlela ezijwayelekile kakhulu.
Kungase kubekwe ikhasimende elikhungathekile kodwa elilula njengeliphuthumayo ngoba ulimi luzwakala luthukuthele. Kungase kubhalwe impendulo eqinisekile esihlokweni sosizo esiphelelwe yisikhathi. Kungase kuqhubeke nokusesha ukuthi isinyathelo esifanele singakanani sokwandisa intengo. Kungase kudalule imininingwane eminingi ye-akhawunti kumbhalo wokuphendula.
Isixazululo akusikho ukuthi “bhala isikhuthazo esingcono” futhi unethemba. Engeza imikhawulo eqinile:
-
Khuphula izinga lapho kuvela khona ulimi lokukhokhisa, lokuphepha, olusemthethweni, noma lokukhansela
-
Dinga izingcaphuno ezivela ezihlokweni zosizo zangaphakathi esifinyezweni sobufakazi
-
Gcina "thumela impendulo" ngemuva kokuvunyelwa ngabantu
-
Ngena ngemvume kuzo zonke izingcingo zamathuluzi kanye nohlaka lokugcina
-
Phinda usebenzise i-test suite yamathikithi angu-30 ngemva kokushintsha kwesicelo, ithuluzi, noma inqubomgomo ngayinye
Ukudla okuwusizo
I-ejenti ewusizo ayidingi ukuzimela okukhulu. Kulesi sibonelo, inani livela ku-loop elawulwayo: funda ithikithi, thola umongo ofanele, uwuhlukanise, ubhale impendulo, bese uyeka ukuze ubuyekezwe. Kulula kakhulu ukwethemba, ukuhlola, nokuthuthukisa lokho kune-ejenti ezama "ukuphatha ukwesekwa" ngesikhuthazo esisodwa esikhulu.
Imibuzo Evame Ukubuzwa
Kalula nje, iyini i-ejenti ye-AI?
I-ejenti ye-AI ngokuyisisekelo iyiluphu ephindaphindayo: thatha okokufaka, unqume isinyathelo esilandelayo, usebenzise ithuluzi, ufunde umphumela, bese uphinda kuze kube yilapho usuqedile. Ingxenye "ye-ejenti" ivela ekusebenzeni nasekubukeni, hhayi nje ekuxoxeni. Ama-ejenti amaningi asebenza ngokuzenzakalela ngokuhlakanipha ngokufinyelela amathuluzi, kanti amanye aziphatha njengomqhubi omncane ongalulama emaphutheni.
Kufanele ngakhe nini i-ejenti ye-AI esikhundleni sokusebenzisa nje i-prompt?
Yakha i-ejenti uma umsebenzi unezinyathelo eziningi, izinguquko zisekelwe emiphumeleni ephakathi, futhi idinga ukusetshenziswa kwamathuluzi okuthembekile (ama-API, izizindalwazi, amathikithi, ukwenziwa kwekhodi). Ama-ejenti nawo ayasiza uma ufuna imiphumela ephindaphindwayo enezivikelo kanye nendlela yokuhlola ukuthi "kuqediwe." Uma impendulo elula esheshayo isebenza, i-ejenti ivame ukuba yi-overhead engadingeki kanye nezindlela zokwehluleka ezengeziwe.
Ngingayakha kanjani i-ejenti ye-AI engabambeki ezinkingeni?
Sebenzisa izimo zokuma okuqinile: izinyathelo eziphezulu, izingcingo eziphezulu zamathuluzi, kanye nokuhlola kokuqeda. Engeza ama-scheme amathuluzi ahlelekile, izikhathi zokuvala, kanye nokuzama kabusha okungeke kuphinde kuzanywe unomphela. Bhala izinqumo kanye nemiphumela yamathuluzi ukuze ubone ukuthi iphambuka kuphi. I-valve yokuphepha evamile ukukhuphula: uma i-ejenti ingaqiniseki noma iphinda amaphutha, kufanele icele usizo kunokuba izenzele.
Iyini ukwakheka okuncane kwe-How to Building Agent AI?
Okungenani udinga iluphu yokulawula epha imodeli umgomo kanye nomongo, icele isenzo esilandelayo, isebenzise ithuluzi uma iceliwe, ifake ukubonwa, bese iphinda. Udinga futhi amathuluzi anezimo eziqinile zokufaka/zokukhipha kanye nokuhlola "okuqediwe". Ngisho neluphu yokugoqa-yodwa ingasebenza kahle uma ugcina isimo sihlanzekile futhi uphoqelela imikhawulo yesinyathelo.
Ngingaklama kanjani ukubizwa kwamathuluzi ukuze kube nokwethenjelwa ekukhiqizeni?
Gcina amathuluzi emancane, athayishiwe, avunyelwe, futhi aqinisekisiwe—gwema ithuluzi elijwayelekile elithi “do_anything”. Khetha ama-schema aqinile (njengokukhipha okuhleliwe/ukushaya umsebenzi) ukuze i-ejenti ingakwazi ukuhambisa ngesandla okokufaka. Engeza uhlu lwabavunyelwe, imikhawulo yesilinganiso, kanye nokuhlolwa kwemvume yomsebenzisi/i-org kusendlalelo samathuluzi. Yakha amathuluzi ukuze aphephe ukuwasebenzisa kabusha uma kungenzeka, usebenzisa amaphethini e-idempotency.
Iyiphi indlela engcono kakhulu yokwengeza inkumbulo ngaphandle kokwenza i-ejenti ibe yimbi kakhulu?
Phatha inkumbulo njengezingxenye ezimbili: isimo sokusebenza kwesikhathi esifushane (izinyathelo zakamuva, uhlelo lwamanje, imikhawulo) kanye nokubuyisa isikhathi eside (izintandokazi, imithetho ezinzile, amadokhumenti afanele). Gcina isikhathi esifushane sihambisana nezifinyezo ezisebenzayo, hhayi imibhalo ephelele. Kwimemori yesikhathi eside, ukubuyisa (ukushumeka + isitolo sevektha/amaphethini e-RAG) kuvame ukushaya "ukugcwala" konke kube umongo futhi kudide imodeli.
Yiliphi iphethini yokuhlela okufanele ngiyisebenzise: uhlu lokuhlola, i-ReAct, noma isisebenzi esiphethe?
Umhleli wohlu lokuhlola muhle kakhulu lapho imisebenzi ibikezelwa futhi ufuna into elula ukuyihlola. Ama-loop esitayela se-ReAct ayakhanya lapho imiphumela yamathuluzi ishintsha lokho okwenzayo ngokulandelayo. Amaphethini omphathi-isisebenzi (njengokuhlukaniswa kwendima yesitayela se-AutoGen) ayasiza lapho imisebenzi ingalinganiswa noma izuze ezindimeni ezihlukile (umcwaningi, umlobi wekhodi, i-QA). Hlela-bese-usebenzisa ngokuhlela kabusha kuyisisekelo esisebenzayo sokugwema izinhlelo ezimbi eziphikelelayo.
Ngingamenza kanjani i-ejenti iphephe uma ingathatha izinyathelo zangempela?
Sebenzisa izimvume ezingenamalungelo amaningi futhi ukhawulele amathuluzi ayingozi ngemuva kokuvunyelwa noma izindlela "zokusebenza ngaphandle kokusebenzisa". Engeza isabelomali kanye nemikhawulo: izinyathelo eziphezulu, ukusetshenziswa okuphezulu, kanye nemikhawulo yokushaya ithuluzi ngomzuzu. Lungisa idatha ebucayi ngaphambi kokubhalisa, bese uhlukanisa i-dev ezindaweni zokukhiqiza. Dinga amafulegi okungaqiniseki noma imibuzo ecacisayo lapho okufakwayo kungacacile, esikhundleni sokuvumela ukuzethemba kuthathe indawo yobufakazi.
Ngingayihlola futhi ngiyihlole kanjani i-ejenti ye-AI ukuze ithuthuke ngokuhamba kwesikhathi?
Yakha i-scenario suite enezindlela ezijabulisayo, amacala onqenqema, ukwehluleka kwamathuluzi, izicelo ezingacacile, kanye nemizamo yokufaka ngokushesha (isitayela se-OWASP). Thola imiphumela efana nempumelelo yomsebenzi, isikhathi sokuqedela, ukululama emaphutheni amathuluzi, kanye nezimangalo ngaphandle kobufakazi. Noma nini lapho ushintsha ama-schema amathuluzi, izikhuthazo, ukubuyisa, noma ukufometha kwememori, sebenzisa kabusha i-suite. Uma ungakwazi ukuyihlola, awukwazi ukuyithumela ngokuthembekile.
Ngingayisebenzisa kanjani i-ejenti ngaphandle kokuphazamisa ukubambezeleka kanye nezindleko?
Iphethini evamile isilawuli esingenasimo esinesitolo sesimo sangaphandle (i-DB/Redis), izinsizakalo zamathuluzi ngemuva kwaso, kanye nokubhalisa/ukuqapha okuqinile (ngokuvamile i-OpenTelemetry). Lawula izindleko nge-caching yokubuyisa, izifinyezo zesimo esihlanganisiwe, amamodeli amancane okuqondisa/ukukhipha, kanye nokunciphisa "ukucabanga okujulile" ezinyathelweni ezinzima kakhulu. Sebenzisa imigqa yemisebenzi emide ukuze ungazigcini izicelo zewebhu zivulekile. Faka njalo inkinobho yokubulala.
Izinkomba
-
Isikhungo Sikazwelonke Sezindinganiso Nobuchwepheshe (i-NIST) - I-NIST AI RMF 1.0 (ukwethembeka kanye nokucaca) - nvlpubs.nist.gov
-
I-OpenAI - Imiphumela Ehlelekile - platform.openai.com
-
I-OpenAI - Umhlahlandlela wokushaya ucingo lomsebenzi - platform.openai.com
-
I-OpenAI - Umhlahlandlela wemikhawulo yamanani - platform.openai.com
-
I-OpenAI - Runs API - platform.openai.com
-
I-OpenAI - Usizo lusebenza ngokubiza - platform.openai.com
-
I-LangChain - Amadokhumenti e-Agents (JavaScript) - docs.langchain.com
-
I-LangChain - Amadokhumenti Amathuluzi (i-Python) - docs.langchain.com
-
I-LangChain - Ukubuka konke kwememori - docs.langchain.com
-
arXiv - Iphepha le-ReAct (isizathu + isenzo) - arxiv.org
-
arXiv - iphepha le-RAG - arxiv.org
-
Umtapo Wokwakha Wezinsizakalo Zewebhu ze-Amazon (AWS) - Ukuphelelwa yisikhathi, ukuzama kabusha, kanye nokubuyela emuva nge-jitter - aws.amazon.com
-
I-OpenTelemetry - Isiqalo sokubuka - opentelemetry.io
-
I-Stripe - Izicelo ze-Idempotent - docs.stripe.com
-
I-Google Cloud - Isu Lokuzama Kabusha (ukubuyela emuva + i-jitter) - docs.cloud.google.com
-
I-OWASP - Eziyi-10 Eziphezulu Zezinhlelo Zokusebenza Zemodeli Yolimi Olukhulu - owasp.org
-
I-OWASP - LLM01 Injection Esheshayo - genai.owasp.org
-
I-LlamaIndex - Isingeniso ku-RAG - developers.llamaindex.ai
-
I-Microsoft - I-Semantic Kernel - learn.microsoft.com
-
I-Microsoft AutoGen - Uhlaka lwama-ejenti amaningi (amadokhumenti) - microsoft.github.io
-
I-CrewAI - Imiqondo yama-ejenti - docs.crevai.com
-
I-Haystack (i-deepset) - Imibhalo ye-Retrievers - docs.haystack.deepset.ai