Riba:
- Ikon kafa amintaccen ginin LLM na girgije wanda baya kiyaye maɓallin API akan abokin ciniki amma yana wucewa ta hanyar wakili na ƙarshe.
- Ikon rubuta haɗe-haɗe masu ƙarfi waɗanda ke haɓaka saurin fahimta tare da yawo kuma a hankali kula da yanayi kamar ƙarewar lokaci, kurakuran hanyar sadarwa da iyakan gudu.
- Ability don rage farashin ta hanyar rage alamar da aka aiko da tambaya game da wajibcin bayanan sirri kafin ya tafi ga girgije.
Kan-na'urar AI yana da ƙarfi amma iyaka. Lokacin da kake son ƙara ainihin "mataimakin taɗi mai wayo," dogon taƙaitaccen rubutu, ko ƙirƙira ƙirƙira zuwa ƙa'ida, kuna buƙatar samfura waɗanda suka fi girma da dacewa da waya. Wannan shine inda girgije AI ya shigo cikin wasa: aikace-aikacenku yana haɗawa da babban ƙirar harshe (LLM) ta hanyar API (Interface Programming Interface - daidaitaccen ƙirar inda software biyu ke aikawa da karɓar bayanai ga juna). A cikin wannan rukunin za mu koyi yadda ake haɗa girgije LLM cikin aikace-aikacen wayar hannu cikin aminci, sauri da kuma tsada. Mahimman mahimmancin mahimmancin tsaro zai kasance akan tsaro: haɗin LLM da ba daidai ba zai iya zubar da maɓallin API ɗin ku kuma ya haifar da lissafin kuɗi na dubban fam.
Tsarin zinariya na gine-gine: kiyaye maɓalli akan abokin ciniki
Kuskure mafi haɗari da za a iya yi a cikin haɗin kai na AI shine shigar da maɓallin API (maɓallin sirrin da ke ba da izinin amfani da sabis ɗin) kai tsaye zuwa lambar aikace-aikacen wayar hannu. Ana sauke aikace-aikacen wayar hannu zuwa na'urar mai amfani kuma ana iya karanta lambar ta hanyar injiniyan juzu'i - nazarin aikace-aikacen da aka haɗa da ganin abin da ke cikinsa. Idan maɓallin ku yana cikin ƙa'idar, wani zai iya cire shi kuma ya yi buƙatu marasa iyaka daga asusunku.
Madaidaicin gine-ginen shine wannan: aikace-aikacen hannu yana aika buƙatun zuwa uwar garken baya (wakilin wakili da kuke sarrafawa); Maɓallin yana zaune ne kawai akan uwar garken; The server goes to the LLM service and returns the response to the application. This middleware also provides speed capping, abuse prevention, and cost control.
kusanci
ina mabudin
Tsaro
Makullin yana cikin aikace-aikacen (FALSE)
A abokin ciniki, jama'a
Yana zubewa, lissafin ya fashe
Makullin yana cikin baya (TRUE)
A kan uwar garken, boye
Amintacce, mai sarrafawa
Tsanaki: Lokacin da ka nemi AI don haɗin haɗin LLM, yana iya samar da misali wanda ke rubuta maɓallin kai tsaye cikin lambar aikace-aikacen don dacewa. Kada ku taɓa ɗaukar wannan rayuwa. Tabbatar cewa kun haɗa jumlar "Kada maɓallin API ya kasance akan abokin ciniki, ku shiga cikin wakili na baya" a cikin hanzari.
Yawo: haɓaka saurin da aka sani
Amsoshin LLM na iya yin tsayi kuma suna ɗaukar daƙiƙa don samarwa gaba ɗaya. Barin mai amfani yana jira akan allo mara kyau shine mummunan kwarewa. Maganin yana gudana - yana nuna kalmar amsa ta kalma, kamar yadda aka samar. Mai amfani yana lura da yadda ake rubuta rubutun, kamar yadda yake cikin ChatGPT; wannan yana ƙaruwa da saurin fahimta da saurin fahimta. Yawo akan wayar hannu yana nufin ƙara guntu (alamu - guntun rubutun da samfurin ya samar) daga uwar garken zuwa mahaɗin yayin da suke isowa. Nemi kwarara a sarari lokacin buga haɗin kai zuwa AI.
Tukwici: Ƙara maɓallin "dakata" a cikin martanin yawo. Ya kamata mai amfani ya iya dakatar da samarwa lokacin da ya sami amsar da yake so; Wannan duka yana inganta ƙwarewa kuma yana rage farashi ta hanyar yanke tsararrun alamun da ba dole ba. A tsakiyar dogon amsar, mai amfani zai iya riga ya sami amsar su.
Kudin, jinkiri da sarrafa kuskure
Cloud LLM yana ɗaukar farashin kuɗi (kudin kowace alama) da farashin lokaci (latency) tare da kowace buƙata. Daban-daban guda uku suna da mahimmanci. Cost: limit prompt and response length, do not send unnecessarily long system instructions, default to small and cheap model if possible. Latency: yi amfani da yawo, saita lokacin ƙarewa, sanar da mai amfani idan cibiyar sadarwa tana jinkirin. Kuskure: katsewar hanyar sadarwa, sabis na iya dawowa 429 (buƙatun da yawa) ko 500 (kuskuren uwar garke); rike kowane ɗayan a hankali, kar a lalata app ɗin. Hakanan, LLM wani lokaci yana ba da amsoshi marasa ma'ana ko kuskure (hallucination); Ƙara Layer na tabbatar da amsar a wurare masu mahimmanci.
uku mini lokuta
Case 1 - Maɓallin Leaked. Farawa ta saka maɓallin buɗe AI kai tsaye cikin ƙa'idar React Native don fita cikin sauri. Makonni uku bayan fitowar manhajar, an canza maɓalli kuma an yi amfani da darajar $2,400 dare ɗaya. Dole ne ƙungiyar ta soke maɓallin kuma ta kafa wakili na baya. Darasi: gajeriyar hanyar da aka ɗauka don dacewa ta zama hanya mafi tsada.
Case 2 - Fitar da aka rage tare da kwarara. Wani app na ilimi ya fara fitar da fasalin Q&A ba tare da yawo ba; masu amfani suna fita bayan daƙiƙa 6 na jiran aiki. Lokacin da aka ƙara kwarara, kalmar farko ta fara bayyana a cikin daƙiƙa 0.8, kuma adadin watsi ya ragu daga 48% zuwa 12%. Samfurin iri ɗaya, gudu ɗaya - kawai bambanci a cikin gabatarwa.
Case 3 - Kula da farashi. Ɗaya daga cikin app yana aika duk tarihin taɗi zuwa samfurin tare da kowane saƙon mai amfani; A cikin doguwar tattaunawa, buƙatu ɗaya ta kai alamun 8,000, wanda ya haɓaka farashin. Ta hanyar aika saƙonnin ƙarshe da taƙaitaccen bayani, ƙungiyar ta rage alamun kowane buƙata da kashi 70%, rage lissafin wata-wata zuwa kashi uku. Darasi: auna abin da kuka aika.
Rauni mai ƙarfi / Ƙarfi mai ƙarfi
Rarraunan faɗakarwa: "Ƙara hira kamar ChatGPT zuwa app na."
Ƙarfafawa mai ƙarfi: "Ƙara mataimakiyar taɗi zuwa aikace-aikacen iOS / Swift na. Gine-gine: aikace-aikacen yana aika buƙatun zuwa nawa na baya, maɓallin API na LLM BA a kan CLIENT ba, yana tafiya ta hanyar wakili. - Amsar ta zo ta gudana, nuna kalma ta kalma - Maɓallin 'Tsaya' ya katse samarwa - Karɓar lokaci, kuskuren hanyar sadarwa, 40en gajeriyar halin da ake ciki da kuma 50000 na ƙarshe aika saƙonnin taɗi na ƙarshe da kuma 5. + taƙaitawa (samar da farashi) Yi bayanin zanen gine-gine da farko, sannan ba abokin ciniki da lambar wakili daban."
Samfura masu kwafi
Samfurin ingantaccen tsarin gine-gine: "Zana haɗin haɗin gwiwar girgije na LLM a cikin aikace-aikacena [dandamali]. Doka: API key kawai a baya. Abokin ciniki -> wakili na -> LLM. A cikin wakili: tabbatarwa, iyakar ƙimar mai amfani, buƙatar shiga. Jerin abokin ciniki da alhakin wakili daban, sannan fitarwa lambar. "
Samfurin yawo: "Ƙara amsa mai yawo zuwa wannan allon taɗi: - Ƙara snippets zuwa kumfa saƙon yayin da suke isowa- Nuna siginan kwamfuta/animation yayin bugawa- Yi maɓallin 'Tsaya' soke rafi- Ajiye ɓangaren rubutu kuma a faɗakar da idan akwai kuskure yayin rafi yana ƙarewa [lambar da ta kasance]"
Samfurin latency mai tsada: "Rage farashi da jinkiri a cikin wannan haɗin gwiwa na LLM: - Ta yaya zan rage alamar da aka aiko (gajartawar tarihi, taƙaitaccen bayani)? - A cikin wane yanayi ƙarami/mai rahusa ya isa? - Ba da shawarar lokacin ƙarewa kuma sake gwada dabara[lamba]"
Samfurin haƙuri na kuskure: "Yi wannan ƙarfin kira na LLM: - Rarrabe hali don babu hanyar sadarwa, ƙarewar lokaci, 429 (iyakar ƙima), 500 (uwar garke) - Saƙon mara fasaha, mai ladabi ga mai amfani- Bayanin tabbatarwa game da haɗarin hasashe a cikin mahimman amsoshi[ladi]"
Kuskuren gama gari
- Saka maɓallin API cikin aikace-aikacen. Mafi tsada da bugu na tsaro; Makullin tabbas yana kwance a ƙarshen baya.
- Ba amfani da kwarara ba. Barin mai amfani yana jiran dogon amsoshi zai kore mai amfani.
- Aika duk tarihin taɗi tare da kowace buƙata. Yana ninka farashin alamar da latency.
- Ketare yanayin kuskure. Idan ba a magance 429/500/lokacin ƙarewa ba aikace-aikacen zai faɗo ko daskare.
- La'akari da amsar LLM daidai ba tare da tambaya ba. Halucination na gaske ne; Ƙara Layer tabbaci a wuri mai mahimmanci.
- Aika bayanan mai amfani zuwa LLM mara amfani. Tambayi ko ana buƙatar bayanan sirri ko ya kamata a rufe su kafin ya tafi ga gajimare.
a takaice
Cloud LLM yana kawo babban ƙarfin da bai dace da na'urar zuwa wayar hannu ba, amma yana buƙatar tsaro da horon farashi. Dokar zinare: Maɓallin API bai taɓa kan abokin ciniki ba, yana wucewa ta hanyar wakili na baya. Gudun gudu yana ƙaruwa da saurin fahimta da riƙewa; Ana goyan bayan maɓallin "tsayawa". An ƙayyade farashin ta hanyar rage alamar da aka aika; Ana samun juriya ta hanyar kula da duk wasu lamurra na kuskure cikin alheri. Amsoshin LLM na iya haɗawa da hallucinations; A cikin wurare masu mahimmanci, tabbatarwa yana da mahimmanci kuma ana duba bayanan sirri kafin aika shi zuwa gajimare.
Aikin aikace-aikace
Nemi abokin ciniki + ƙirar wakili na baya daga AI ta amfani da "Tsarin tsarin gine-gine" don fasalin "takaitaccen rubutu" ko "tattaunawa". Tabbatar da cewa maɓallin API kawai yana zaune a bayan baya a ƙirar da aka ƙirƙira. Sannan cire aƙalla hanyoyi biyu don rage alamar da aka aiko tare da tsarin “Cost-delay pattern” kuma rubuta saƙon ladabi don nunawa ga mai amfani don yanayin kuskure (misali 429).
jerin abubuwan dubawa
- [ ] Na tabbatar da cewa maɓallin API yana zaune a bayan baya ba akan abokin ciniki ba
- [ ] Na sanya martanin yawo kuma na ƙara maɓallin 'dakata'
- [] Na magance lokacin ƙarewa, kuskuren hanyar sadarwa, yanayi 429 da 500
- [ ] Na rage alamar da aka ƙaddamar tare da gajarta/taƙaice da ta gabata
- [ ] Na yi la'akari da inganci game da haɗarin hasashe a cikin amsar LLM
- [ ] Na duba wajibcin / rufe bayanan sirri kafin in je gajimare