- Ama-harnesses e-AI ahlinzeka ngengqalasizinda ebalulekile—kufaka phakathi ama-sandbox kanye nesimo esiqinile—esiguqula ama-LLM angavuthiwe abe ama-ejenti asebenzayo.
- UJev wethula imodeli yesinqumo se-System One enikeza enye indlela esheshayo nengabizi kakhulu kune-LLM yendabuko yokuhlukaniswa okuhlelekile.
- Ukuhlanganisa i-Jev ne-LangChain kanye ne-LangGraph kuvumela onjiniyela ukuthi bakhe izinhlelo 'zokukhiqiza, hhayi unkulunkulu' ngokuhlanganisa ukucabanga kusukela ekwenzeni izinqumo.
Uma sikhuluma ngama-ejenti e-AI, abantu abaningi bagxila ebuchosheni—i-Large Language Model. Kodwa nansi inkinga: imodeli ngokwayo imane nje iyisibikezeli sombhalo esimangalisayo. Ukuze wenze izinto emhlabeni wangempela, udinga lokho okwaziwa ngokuthi i- harness . Cabanga nge-harness njengomzimba namathuluzi avumela ubuchopho ukuthi busebenzisane nomhlaba, bulawule konke kusukela enkumbulweni kuya ekusebenzeni kwangempela kwekhodi.
Muva nje, umdlalo ushintshile ngokufika kwe -Jev evela ku-TypeSafe AI . Esikhundleni sokuphatha zonke izinqumo ezincane njengempikiswano yefilosofi edinga i-LLM enkulu, i-Jev iphatha izinketho ezihlelekile nezisheshayo. Ngokuhlanganisa lokhu ne -LangChain ne-LangGraph , abathuthukisi ekugcineni bangasuka ekwakhiweni kwezakhiwo okubizayo, okuhamba kancane baye kokuthile okuthambile, okubi, nokukulungele ukukhiqiza.
Ukuqonda Ukwakheka Kwe-Agent Harness
Uma ungeyena oqeqesha imodeli, empeleni ungunjiniyela we-harness. I-harness yiqoqo lonke lekhodi, ukucushwa, kanye ne-logic elihlanganisa imodeli. Yilokho okuguqula i-LLM eluhlaza ibe yinjini yokusebenza ngokunikeza isimo, ukufinyelela kwamathuluzi, kanye nemikhawulo eqinile. Ngaphandle kwalolu ngqimba, imodeli ayikwazi ukukhumbula ukuthi yenzeni emizuzwini emibili edlule noma empeleni ichofoze inkinobho kuwebhusayithi.
Esinye sezingcezu ezibaluleke kakhulu ze-harness yi- filesystem . Ngokunika i-ejenti indawo yokusebenza, ingalayisha idatha engangeni efasiteleni lomongo futhi igcine isimo esiqinile kuzo zonke izikhathi ezahlukene. Lokhu kuvimbela inkinga evamile "yokubola komongo," lapho i-ejenti ilahlekelwa indaba njengoba ingxoxo iba yinde kakhulu. Ama-harnesses athuthukile asebenzisa ukucindezela kanye nokukhipha i-tool-call ukuze agcine ulwazi olufanele kakhulu luphambili futhi luphakathi ngenkathi egcina okunye kudiski.
Ukuze ama-ejenti azimele ngempela, ama-harnesses avame ukufaka ithuluzi le-bash kanye nama-sandbox avikelekile . Esikhundleni sokubhala ikhodi eqinile kuwo wonke amathuluzi angakhona, unikeza i-ejenti i-terminal. Lokhu kuyivumela ukuthi ibhale izikripthi zayo ukuze ixazulule izinkinga ngokushesha. Kodwa-ke, ukusebenzisa ikhodi ekhiqizwe yi-ejenti kuyingozi, yingakho izindawo zokusebenzisa ezihlukanisiwe zingaxoxiswani nganoma yiluphi uhlelo lwebanga lokukhiqiza.
Faka uJev: Imodeli Yesinqumo Sesistimu Eyodwa
Iningi labathuthukisi bebesebenzisa ama-LLM asemngceleni kukho konke, okufana nokusebenzisa i-supercar ukushayela amamitha ayishumi ukuya ebhokisini leposi. Kuyihaba kakhulu. UJev uyakushintsha lokhu ngokusebenza njengemodeli yeSistimu Yokuqala . Akwenzi iphrozi; wenza izinqumo ezibhalwe ngomshini . Kungakhathaliseki ukuthi u-yebo/cha (uNoul), ukukhetha ohlwini, noma amaphuzu ezinombolo, uJev unikeza izimpendulo ngamathuba ahlobene.
Ukukhuphuka kokusebenza kukhulu impela. Sikhuluma ngesivinini sokuphetha esifinyelela ku-200x ngokushesha kanye nezindleko eziphansi ngo-400x kune-LLM zendabuko zokuhlukaniswa. Ngenxa yokuthi i-Jev ingahlola imibuzo eminingi ngesikhathi esifanayo engxenyeni efanayo yesimo, ilungele imisebenzi yokuhlunga, njengokunquma ukuthi ithikithi lokusekela amakhasimende liyaphuthuma noma ukuhlonza i-PII kudokhumenti yezomthetho.
Ukusebenzisa i-Jev ngaphakathi kwe-LangChain Ecosystem
Ukuhlanganisa i-Jev ku-LangChain kulula nge- TypeSafeClassifier . Esikhundleni sempendulo yengxoxo, uthola ukuhlukaniswa okuhlelekile. Lokhu kukuvumela ukuthi wakhe ungqimba lwesinqumo oluhlala phambi kwe-ejenti yakho eyinhloko. Isibonelo, ungasebenzisa i -Request Router ehlaziya umbuzo bese ithumela imisebenzi elula kumodeli eshibhile njenge-GPT-4o-mini kanye nemisebenzi yokwakha eyinkimbinkimbi ku-powerhouse efana ne-GPT-4o.
Esinye isimo sokusetshenziswa esibulalayo yi- Tool Gate . Ngaphambi kokuba i-ejenti ikhiphe umyalo ongaba yingozi—njengokususa i-database—i-harness ingasebenzisa i-Jev ukuhlola izinga lobungozi . Uma ingozi iphezulu kakhulu, i-harness ingavimba ucingo noma ibangele ukwanda komuntu ngaphakathi kwe-loop. Lokhu kudala inethi yokuphepha ethembeke kakhulu kunokumane ucele i-LLM ukuthi “iqaphele” esicelweni sesistimu.
Ukuhlanganiswa neLangGraph kanye noKuhlukaniswa koBunhloli
Nakuba i-Jev iphatha izinqumo, i-LangGraph inikeza uhlelo lokuqondisa . Ikuvumela ukuthi uchaze ukugeleza kwe-ejenti njengegrafu yama-node kanye nemiphetho. Lokhu kusho ukuthi isahlulelo sivela kumodeli, kodwa ukugeleza kuhlala kukhodi , okwenza uhlelo lubonakale futhi luhlolwe. Ungasebenzisa izindawo zokuhlola ukuqinisekisa ukuthi uma inqubo yehluleka, ingaqhubeka lapho ishiye khona ngaphandle kokuchitha amathokheni.
Lolu shintsho yilokho ochwepheshe abakubiza ngokuthi i- Great Unbundling of Intelligence . Esikhundleni sokuba imodeli eyodwa enkulu yenze konke, siya kumodeli "eshibhile ngokuzenzakalelayo, engenamkhawulo ngaphandle". Esimweni sokuzenzakalela kwesiphequluli, i-Jev ingakhetha ngokushesha ukuthi iyiphi into okufanele ichofoze; uma amaphuzu okuzethemba ehla ngaphansi komkhawulo othile (isb., 0.7), uhlelo lukhulisa ngokuzenzakalelayo umsebenzi luye ku-LLM enekhono kakhulu lokucabanga okujulile.
Ukusebenzisana phakathi kwe-harness eqinile, imodeli yesinqumo esisheshayo njenge-Jev, kanye nesikhathi sokusebenza esihlelekile njenge-LangGraph kuvumela ukuzimela okude . Ngokusebenzisa amaphethini afana ne-Ralph Loop—efaka kabusha izikhuthazo zokuqala kumafasitela omongo ahlanzekile—ama-ejenti angasebenza kumaphrojekthi ayinkimbinkimbi, anezinyathelo eziningi ngaphandle kokulahlekelwa ukuhambisana. Umgomo wokugcina uhlelo lapho imodeli inikeza ubuhlakani, futhi i-harness yenza lobo buhlakani bube wusizo futhi buphephile ekusetshenzisweni kwangempela.
