Cishe usujwayele ukucabanga ngobuhlakani bokwenziwa njengaleyo chatbot ekubhalela izinkondlo noma efingqa ama-imeyili, kodwa kukhona umkhuba omusha odlula amagama ngokuphelele. I-TypeSafe AI yethule i-Jev , ithuluzi eliphula isikhunta ngoba alibhali umusho owodwa; umgomo wayo ukuletha izinqumo zobuchwepheshe nezingenzeka ezingacutshungulwa yinoma yiluphi uhlelo lwekhompyutha ngokushesha ngaphandle kokuphazamiseka.
Cabanga ukuthi unesistimu ephatha izinkulungwane zezicelo, futhi awudingi i-AI ukuchaza ukuthi kungani ithikithi liphuthuma; udinga nje ukuthi ithi "kuphuthuma: yebo." UJev uzibeka lapho ngqo, kulelo gebe phakathi kwekhodi eqinile namamodeli olimi oluhamba kancane, enikeza lokho abakubiza ngokuthi amamodeli eSistimu One , aphefumulelwe yingqondo kaDaniel Kahneman yokucabanga okusheshayo nokunembile.
Iyini ngempela i-Jev AI?
Ngokungafani nama-giants afana ne-GPT noma i-Claude, akhipha amathokheni ngalinye ngalinye ukwakha isigaba, i-Jev isebenza njenge- smart function call . Uyidlulisela esimweni (esingaba yi-JSON, umbhalo, noma uhlu) kanye nesethi yemibuzo echazwe kusengaphambili, bese imodeli ibuyisa amanani ahlelekile. Lokhu kusho ukuthi azikho amaphutha ohlobo noma izinkinga zokufometha, njengoba impendulo iboshelwe ezinkethweni ozikhethile ngaphambilini.
Ukwakheka kwale modeli kusebenzisa indlela ethile yokuqeqesha ebizwa ngokuthi i-Reinforcement Learning for Calibrated Decisions (RLCD) . Umgomo lapha awukona ukuthi i-AI ithandeke noma ikholise, kodwa kunalokho ukuthi amathuba ayo abe qotho futhi alinganiswe. Uma uJev ethi uqiniseka ngo-90% empendulweni, leso sibalo kufanele sibonise ukunemba kwangempela—into ebalulekile uma ufuna isofthiwe isebenze ngokuzimela ngaphandle kokuthi umuntu abukeze isinyathelo ngasinye.
Amathuluzi amathathu ayisisekelo: izinto zokuqala
Ukuze izinto zihlale zilula futhi zisebenza kahle, i-Jev's API isekelwe ezinhlotsheni ezintathu zemibuzo ezivumela ukuthi ihlanganise cishe noma yisiphi isidingo sokuhlukaniswa:
- Ukukhetha: Ikuvumela ukuthi ukhethe inketho eyodwa eqenjini lamathuba angafika ku-255. Ayikuniki nje kuphela inketho yokuwina, kodwa futhi imephu yamathuba azo zonke izinketho kanye nesilinganiso sokuzethemba. Ifaka inketho "enye" ​​yokuvimbela imodeli ukuthi ingaphoqi impendulo engalungile.
- Umphumela: Hlola okuthile esikalini esiphothuliwe (kusukela kumazinga ama-2 kuya kwayi-10). Into ethokozisayo ukuthi ibuyisa amanani edesimali (njengo-1.4), okuvumela ukuthi uthwebule... izici ezihlukile phakathi kwezigaba ukuthi inombolo ephelele izoyishaya indiva.
- I-Noul (ukuhlolwa kwe-Boolean): Kuyinguqulo elula ka-yebo noma cha, okubuyisela amathuba phakathi kuka-0 no-1.
Isici esinamandla kakhulu ukuthi uJev angaphendula yonke le mibuzo ngesinyathelo esisodwa, esifanayo . Uma udinga ukwazi isigaba somyalezo, izinga lentukuthelo yekhasimende, nokuthi ngabe licela ukubuyiselwa imali, uJev ukunika zonke izimpendulo ezintathu ngesikhathi esisodwa, ngakho-ke ukwengeza eminye imibuzo akuthinti kangako isikhathi sokuphendula.
Ukusebenza, izindleko, kanye nombhalo omncane
Uma ubheka izinombolo, i-TypeSafe AI inesivinini esimangalisayo, ngama-latency asukela kuma -millisecond angu-70 kuya kwangu-500 . Uma kuqhathaniswa nama-LLM asemngceleni, angathatha imizuzwana ukubhala impendulo, i-Jev ishesha ngendlela exakile. Ngaphezu kwalokho, intengo iyinzuzo ebalulekile: ibiza amathokheni okufaka angu-$0,042 ngesigidi , futhi okukhiphayo kumahhala ngokuphelele, njengoba kungakhiqizi umbhalo omusha.
Kodwa-ke, lezi zibalo kufanele zithathwe ngosawoti omncane. Ama-benchmark amaningi enziwe yithimba le-TypeSafe uqobo, okungase kuveze ukucwasa okuthile. Noma kunjalo, isithembiso samaphutha e-schema awekho siyabonakala: akunakwenzeka ngokwezibalo ukuthi i-Jev ibuyisele ifomethi ongayichazanga, okususa isidingo sokuqinisekiswa okuyinkimbinkimbi kukhodi.
I-Jev AI ikhanya kuphi futhi iwaphi amaphutha?
I-Jev iyinkosi uma kufanele wenze izinqumo eziphindaphindwayo nezinkulu ngedatha engahlelekile. Ilungele ukuqondisa ama-lead , ukulinganisela okuqukethwe, ukulinganisa ukwesekwa kobuchwepheshe, noma ngisho nokudala izivikelo zokuqapha ukuphuma kwamanye amamodeli e-AI amakhulu. Iwusizo kakhulu nasekuzulazuleni kwewebhu, lapho inganquma khona ukuthi iyiphi into okufanele ichofoze ekhasini ngaphandle kokudinga incazelo yesikrini.
Kodwa-ke, ungazami ukuyisebenzisa kukho konke. I-Jev ayifaneleki ukuxoxa, ukuhlela, noma ukuchaza ukucabanga kwayo. Inemikhawulo ecacile ngokubalwa kwezibalo kanye nezinhlamvu, futhi ayiphathi kahle izinsuku noma amanani e-hexadecimal. Ngaphezu kwalokho, ayinalo ulwazi lomhlaba; yazi kuphela lokho okufakayo esimweni sombuzo. Uma udinga ukucabanga okuyinkimbinkimbi noma ukukhiqiza okudaliwe, kufanele unamathele kumodeli yeSistimu 2 efana neClaude noma i-GPT.
Ukuhlanganiswa ohlelweni lwentuthuko
Kwabahleli bezinhlelo, ukuhlanganiswa kulula kakhulu. Ama-SDK asemthethweni akhona eJavaScript, TypeScript, kanye nePython , futhi yize engekho e-PHP, isicelo se-POST esilula yikho konke okudingekayo. Ngaphezu kwalokho, i-Jev isivele itholakala ku-AI Gateway kaVercel, okwenza kube lula kakhulu ukuphathwa kwezihluthulelo kanye nokukhokhisa uma usuvele usebenzisa leyo ndawo.
Iphethini ehlakaniphile kakhulu yokusebenzisa i-Jev ukudala i- cascade yezinqumo . Okokuqala, i-Jev ihlunga futhi ihlukanise isicelo ngentengo ephansi futhi ngokushesha; bese kuba ikhodi eqondile ephatha amacala alula, futhi kuphela ezimweni lapho i-Jev ibonisa ukuzethemba okuphansi lapho isicelo sikhuphukela kumodeli yokucabanga ebizayo noma kumqhubi womuntu. Lokhu kuthuthukisa isabelomali futhi kunciphisa ukubambezeleka kwesistimu iyonke.
Leli thuluzi lichaza kabusha ubudlelwano phakathi kwe-AI nesofthiwe ngokwenza ubuhlakani bube yingxenye yengqalasizinda kunokuba kube yinto exhumana nabantu. Ngokususa ukukhiqizwa kombhalo nokugxila ekulinganisweni okungenzeka, i-Jev ivumela imishini ukuthi yenze izinqumo ezisheshayo nezithembekile ngezindleko ezincane, ivula umnyango wezenzo ezenzakalelayo ezazibiza kakhulu noma zihamba kancane ngaphambilini.