Yeqa uye kokuyinhloko
Izifundo ze-AI

Yakha futhi Uthuthukise Imisebenzi ye-AI nge-jev-align

Funda ukufaka i-jev-align, ukulungiselela umsebenzi wokuhlukanisa noma wokunikeza amaphuzu we-AI, ukumaka izibonelo ezingaqinisekile, ukubuyekeza iziphakamiso ze-GEPA, ukukhetha abahlinzeki bamamodeli, nokuqoqa idatha yokukhiqiza ukuze ithuthukiswe kamuva.

Cishe imizuzu engu-24 yokufunda
Yakha futhi Uthuthukise Imisebenzi ye-AI nge-jev-align

Yini i-jev-align?

jev-align iyithuluzi lokuhlola elisemgqeni womyalo elisahlolwa, elivela kwa-Sutro, lokwakha ama-AI Function nge-Jev ye-TypeSafe. Ihlola idathasethi, ikhombe izibonelo ezingaqinisekile, ikucele ukuthi uzilebule, bese isebenzisa i-GEPA ukuphakamisa incazelo yomsebenzi ethuthukisiwe.

Leli thuluzi lenzelwe ukuhamba komsebenzi okuphindaphindwayo. Ungakha i-AI Function kudatha yasendaweni, ubuyekeze izinguquko eziphakanyisiwe, ulondoloze inqubekelaphambili yakho, usebenzise umsebenzi kuhlelo lokusebenza, bese kamuva uwuthuthukisa ngezibonelo eziqoqwe ekukhiqizweni.

Yonke ilebula ivela kuwe. Isikolo esiphezulu sokuqeqesha asibangeli ukuthi i-jev-align yamukele isiphakamiso ngokuzenzakalelayo.

Izici ezibalulekile

  • Ukulebula okuqhutshwa ukungaqiniseki: I-CLI ibeka phambili imigqa engaqondakali, iphinde ifake nesampula yokuhlola ekhethwe ngokungahleliwe.
  • Ukuthuthukisa okulawulwa ngumuntu: Nguwe onqumayo ukuthi wamukele noma wenqabe incazelo ngayinye ephakanyisiwe.
  • Izinhlobo zemisebenzi eziningi: Yakha imisebenzi ye-binary, multiclass, multilabel, noma yokukala ehleliwe.
  • Ukutholwa kwamadathasethi endawo: Ukusetha okuqondisiwe kuthola amafayela e-CSV, e-Parquet, kanye ne-JSONL.
  • Abahlinzeki bamamodeli abaguquguqukayo: Sebenzisa i-Jev nge-TypeSafe AI, Vercel AI Gateway, noma Cloudflare Workers AI.
  • Amamodeli okucabanga ahlukene: Lungiselela i-OpenAI, Anthropic, Gemini, noma omunye umhlinzeki ohambisana ne-LiteLLM we-GEPA.
  • Ukuqhutshwa okuqhubekayo: Qhubeka nezikhathi zangaphambilini zokuthuthukisa futhi uhlole ama-AI Function alondoloziwe.
  • Ukuqoqwa kokukhiqizwa: Qoqa izibonelo eziwusizo zohlelo lokusebenza bese uzilebula phakathi nomjikelezo wokuthuthukisa ozayo.

Qonda izinhlobo zemisebenzi ezisekelwayo

Ngaphambi kokuqala, khetha ifomethi yokuphumayo ehambisana nenkinga yakho:

  • Binary: Ibuyisa True noma False.
  • Multiclass: Ibuyisa ilebula eyodwa ncamashi kusethi engaguquki.
  • Multilabel: Ibuyisa amalebula ayiziro noma amaningi kusethi engaguquki.
  • Score: Ibuyisa izinga elilodwa kusikali esihleleke ngokulandelana.

Faka i-jev-align

Okudingekayo ngaphambi kokuqala

Udinga i-Python 3.11 noma entsha. Udinga futhi imininingwane yokufinyelela yomhlinzeki we-Jev kanye nomhlinzeki wemodeli yokucabanga osekelwayo.

Faka nge-uv

Ukufakwa okunconyiwe kokuqala ngokushesha kusebenzisa i-uv:

uv tool install jev-align

Faka nge-pip

Uma ungasebenzisi i-uv, faka iphakheji nge-pip:

pip install jev-align

Lungiselela imininingwane yokufinyelela

Endleleni ejwayelekile ye-TypeSafe AI, thumela ukhiye we-TypeSafe API njengokuguquguquka kwemvelo. I-GEPA nayo idinga umhlinzeki wemodeli yokucabanga, njenge-OpenAI, Anthropic, noma Gemini.

export TYPESAFE_API_KEY="..."
export OPENAI_API_KEY="..."

Ungasebenzisa i-ANTHROPIC_API_KEY noma i-GEMINI_API_KEY esikhundleni sikakhiye we-OpenAI. I-Vercel AI Gateway ne-Cloudflare Workers AI zingangena esikhundleni se-TypeSafe AI njengomhlinzeki we-Jev; lezi zindlela azidingi TYPESAFE_API_KEY.

Qalisa ukusetha okuqondisiwe

Qalisa i-CLI nganoma yimuphi wale miyalo:

jeva

noma:

jev-align

Ukusetha okuqondisiwe kuthola amadathasethi endawo e-CSV, e-Parquet, kanye ne-JSONL futhi kufaka izibonelo ezintathu ezilungele ukusetshenziswa. Landela imiyalelo ukuze ukhethe idathasethi, uchaze umsebenzi, ukhethe izinkambu ezifanele, futhi ulungiselele abahlinzeki.

Izilawuli zokuzulazula ze-CLI

  • Sebenzisa okhiye bemicibisholo no-Enter ukuzulazula kumamenyu.
  • Cindezela u-b ukuze ubuyele kulebula yangaphambilini.
  • Faka u-/back ukuze uphume ekwazisweni kwesizathu.
  • Cindezela i-Space ukuze ushintshe izinketho emsebenzini we-multilabel.

Qalisa umsebenzi wokuthuthukisa usebenzisa amafulegi

Ungakwazi ukweqa izingxenye zokuhamba komsebenzi okuqondisiwe ngokunikeza izinketho zomugqa womyalo. Isibonelo esilandelayo sakha umsebenzi we-binary onquma ukuthi okuthunyelwe kumayelana nezindiza yini:

jeva optimize posts.csv \
  --question "Is the post related to aviation?" \
  --column title \
  --column text \
  --pool-size 1000

Izinketho eziphindaphindiwe ze---column zitshela i-jev-align ukuthi ihlole izinkambu ze-title ne-text. Inketho ethi --pool-size 1000 ikhawulela idamu lokusebenza emigqeni engu-1,000.

Ngokuzenzakalelayo, i-jev-align isebenzisa imigqa yokuqala engu-1,000, noma yonke idathasethi uma inemigqa engaphansi kwaleyo. Ungahlanganisa zonke izinkambu noma ukhethe amakholomu athile.

Lungiselela amakilasi namazinga esikolo

Emisebenzini ye-multiclass ne-multilabel, nikeza izinketho eziphindaphindiwe ze---class "NAME=DESCRIPTION". Emisebenzini yokukala, nikeza izinketho eziphindaphindiwe ze---score-level. Njengoba amafulegi atholakalayo engashintsha, bheka inkomba eyakhelwe ngaphakathi:

jeva optimize --help

Indlela umjikelezo ngamunye wokuthuthukisa osebenza ngayo

  1. Hlola idathasethi: I-jev-align isebenzisa umsebenzi wamanje kudatha elungiselelwe bese ikala ukungaqiniseki.
  2. Khetha izibonelo: Ikhetha imigqa engaqondakali bese yengeza nesampula yokuhlola ekhethwe ngokungahleliwe.
  3. Qoqa amalebula: Ulebula izibonelo ezikhethiwe futhi ungakwazi ukunikeza nezizathu ngokuzikhethela.
  4. Qalisa i-GEPA: I-GEPA isebenzisa impendulo eqoqiwe ukuphakamisa incazelo ethuthukisiwe.
  5. Buyekeza isiphakamiso: I-CLI ibonisa isikolo, ushintsho ekuqinisekeni, kanye nomehluko wencazelo ephakanyisiwe.
  6. Khetha isenzo esilandelayo: Ungamukela, wenqabe, ubuyele emuva, noma ume bese uqhubeka kamuva.

Lesi sinyathelo sokubuyekeza sibalulekile: ukuthuthukiswa kwamaphuzu kuhlala kuyiziphakamiso uze ukwamukele ngokusobala.

Lungisa izilungiselelo ezithuthukile zokwenza kahle

Imenyu eqondisiwe ethi Advanced iveza izilawuli eziningana:

  • Khetha izichasiselo zokuqeqesha ezingu-5, 10, 15, noma 20 emzuliswaneni ngamunye.
  • Nika amandla isethi yokuhlola engu-20% egcinelwe ukuhlolwa ngokuzikhethela.
  • Setha isabelomali sezingcingo zemethrikhi se-GEPA, esizenzakalelayo sibe ngu-300.

Imizuliswano emincane yezichasiselo ingenza umjikelezo ngamunye wokubuyekeza usheshe. Isethi egcinelwe ukuhlolwa inganikeza isignali yokuhlola ehlukile ezibonelweni ezisetshenziselwa ukwenza kahle. Isabelomali sezingcingo zemethrikhi silawula ukuthi i-GEPA ingakanani umsebenzi engawenza ngesikhathi sokuqalisa.

Khetha umhlinzeki we-Jev

Ukusetha okuqondisiwe kuthola abahlinzeki abalungiselelwe futhi kukuvumela ukhethe phakathi kwe-TypeSafe AI, i-Vercel AI Gateway, ne-Cloudflare Workers AI. Umhlinzeki okhethiwe ugcinwa ne-AI Function, ngakho izingcingo zesikhathi sokusebenza kamuva zisebenzisa umzila ofanayo.

Sebenzisa i-Vercel AI Gateway

export AI_GATEWAY_API_KEY="..."
jeva optimize data.csv --question "Is this relevant?" --column text \
  --backend vercel

Sebenzisa i-Cloudflare Workers AI

export CLOUDFLARE_ACCOUNT_ID="..."
export CLOUDFLARE_API_TOKEN="..."
jeva optimize data.csv --question "Is this relevant?" --column text \
  --backend cloudflare

Ayikho kule mizila edinga i-TYPESAFE_API_KEY.

Lungiselela imodeli yokubonisa ye-GEPA

Imodeli yokubonisa ye-GEPA ihlukile kumodeli ye-Jev ehlola imigqa yedathasethi. Amamodeli e-OpenAI, Anthropic, ne-Gemini atholwa ngokuzenzakalelayo. Ungasebenzisa futhi noma yimuphi umhlinzeki we-LiteLLM, okuhlanganisa i-Fireworks, i-vLLM yasendaweni, namanye ama-endpoint ahambisana ne-OpenAI.

Ngokwesibonelo, lungiselela i-endpoint ye-vLLM esingathwe bese ukhetha imodeli yayo ngokusobala:

export HOSTED_VLLM_API_BASE="http://localhost:8000/v1"
jeva optimize data.csv --question "Is this relevant?" --column text \
  --reflection-model "hosted_vllm/Qwen/Qwen3-8B"

Lokhu kwehlukanisa kukuvumela ukhethe umhlinzeki oyedwa wokuhlola idathasethi yesikhathi sokusebenza nomunye womsebenzi wokwenza kahle we-GEPA.

Phatha futhi uqhubeke nama-AI Function agciniwe

Bhala uhlu lwemisebenzi yakho egciniwe usebenzisa:

jeva functions

Ukuze uqhubeke nokuqalisa okuthile ngokuqondile, nikeza uhla lwemibhalo olugciniwe:

jeva optimize --resume .jev-align/runs/<run-id>

Ukuqhubeka kuyasiza uma ufuna ukuqhubeka nokulebula, uvakashele kabusha i-function egciniwe, noma ucubungule izibonelo eziqoqwe ngemva kokushicilela.

Sebenzisa i-AI Function ku-Python

Layisha i-function egciniwe usebenzisa i-AIFunction.load. Nika amandla ukuqoqwa ukuze i-function igcine izibonelo eziwusizo zokukhiqiza ukuze zibuyekezwe esikhathini esizayo:

from jev_align import AIFunction

is_aviation = AIFunction.load(
    ".jev-align/runs/<run-id>",
    capture=True,
)

prediction = is_aviation(
    title="Airport expansion",
    text="A new runway opens next year.",
)

Izimpikiswano zamagama kufanele zihambisane nezinkambu ezilindelwe yi-function egciniwe. Kulesi sibonelo, i-function ihlola isihloko nomzimba wombhalo.

Qhubeka ufunde kudatha yokukhiqiza

Uma usebenzisa i-capture=True, uhlelo lwakho lungagcina izibonelo eziwusizo ekwenzeni kahle kwesikhathi esizayo. Buyela ku-CLI uma usulungele ukuzibuyekeza:

jeva functions

Khetha noma uqhubeke ne-AI Function efanele, lebula izibonelo eziqoqiwe, bese uvumela i-GEPA iphakamise incazelo elandelayo. Inqubo efanayo yokubuyekezwa komuntu iyasebenza: hlola amamethrikhi nomehluko ngaphambi kokwamukela ushintsho.

Amathiphu okusebenza athuthukile

  • Khetha amakholomu agxile: Faka kuphela izinkambu ezinikeza ubufakazi bomsebenzi, kunokusebenzisa ngokuzenzakalelayo zonke izinkambu.
  • Bhala umbuzo onembile: Umngcele ocacile wesinqumo wenza ukuhlukanisa okubili kube lula ukukulebula ngokungaguquguquki.
  • Chaza amalebula angaguquki ngokucophelela: Emisebenzini enamakilasi amaningi namalebula amaningi, sebenzisa izincazelo zamakilasi ukuhlukanisa amalebula angase agqagqane ngenye indlela.
  • Faka izizathu lapho kuwusizo: Izizathu ozikhethela zona zinganikeza i-GEPA umongo owengeziwe wokuthi kungani isibonelo kufanele sithole ilebula ethile.
  • Sebenzisa isampula yokuhlola: Izibonelo ezingacacile zibalulekile, kodwa izibonelo ezingahleliwe zisiza ukuhlola ukuziphatha ngaphandle kwesethi egxile ekungaqinisekini.
  • Buyekeza okungaphezu kwamaphuzu: Hlola izinguquko zokuqiniseka nomehluko wencazelo ophakanyisiwe ngaphambi kokwamukela ukuphindaphindwa.
  • Qhubeka esikhundleni sokuqala phansi: Ukuqalisa okugcinwe kugcina amalebula aqoqiwe futhi kwenza kube nokwenzeka ukuqhubeka nokuthuthukisa.
  • Hlukanisa izinqumo zabahlinzeki: Khetha i-backend ye-Jev ngezidingo zokuhlola nesikhathi sokusebenza, bese ulungiselela imodeli yokubonisa ngokuzimela ye-GEPA.

Isiphetho

jev-align inikeza inqubo enomuntu osesigungwini sokulawula yokudala nokuthuthukisa ama-AI Function e-Jev. Qala ngedathasethi yasendaweni, lebula izibonelo ezingaqinisekile, buyekeza izinguquko zencazelo eziphakanyiswe yi-GEPA, bese ugcina umphumela. Ngokunika amandla ukuqoqwa kokukhiqiza nokuqhubeka nokuqalisa okufanayo kamuva, ungaqhubeka nokuthuthukisa i-function ngezibonelo zangempela zohlelo lokusebenza, kuyilapho ugcina ukulawula kuzo zonke izinguqulo ezamukelwayo.