Modern yaz\u0131l\u0131m geli\u015ftirmenin ve \u00f6zellikle yapay zeka destekli sistemlerin kalbinde, farkl\u0131 servislerin ve ara\u00e7lar\u0131n birbiriyle etkile\u015fim kurmas\u0131 yatar. Ancak bu etkile\u015fimler, \u00f6zellikle ger\u00e7ek zamanl\u0131 veya “ak\u0131\u015f” (streaming) ba\u011flam\u0131nda ger\u00e7ekle\u015fti\u011finde, karma\u015f\u0131kl\u0131k h\u0131zla artabilir. Peki, ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 y\u00f6netirken performans sorunlar\u0131, hata ay\u0131klama kabuslar\u0131 ve durum y\u00f6netimi zorluklar\u0131 aras\u0131nda kaybolmadan, ak\u0131l sa\u011fl\u0131\u011f\u0131n\u0131z\u0131 koruyarak nas\u0131l ba\u015far\u0131l\u0131 olabilirsiniz? Bu makale, bu zorlu\u011fun \u00fcstesinden gelmek i\u00e7in stratejiler, pratik \u00f6rnekler ve ileri d\u00fczey teknikler sunarak, sizi s\u0131f\u0131rdan uzmanl\u0131k seviyesine ta\u015f\u0131yacak kapsaml\u0131 bir rehber niteli\u011findedir.<\/p>\n
Yapay zeka modellerinin, \u00f6zellikle B\u00fcy\u00fck Dil Modellerinin (LLM’ler), yetenekleri giderek art\u0131yor. Ancak bu modellerin kendi ba\u015flar\u0131na yapabilecekleri \u015feyler s\u0131n\u0131rl\u0131d\u0131r; ger\u00e7ek d\u00fcnya etkile\u015fimi i\u00e7in harici “ara\u00e7lara” (tools) veya API’lere ihtiya\u00e7 duyarlar. Bir ara\u00e7 \u00e7a\u011fr\u0131s\u0131 (tool call), bir yapay zeka modelinin belirli bir g\u00f6revi yerine getirmek i\u00e7in harici bir servisi veya fonksiyonu \u00e7a\u011f\u0131rmas\u0131 anlam\u0131na gelir. \u00d6rne\u011fin, bir hava durumu API’sini sorgulamak, bir e-posta g\u00f6ndermek veya bir veritaban\u0131ndan bilgi almak gibi i\u015flemler ara\u00e7 \u00e7a\u011fr\u0131lar\u0131 arac\u0131l\u0131\u011f\u0131yla yap\u0131l\u0131r. Bu \u00e7a\u011fr\u0131lar, modelin yeteneklerini geni\u015fleterek, onu daha dinamik ve kullan\u0131\u015fl\u0131 hale getirir.<\/p>\n
“Ak\u0131\u015f” (streaming) ba\u011flam\u0131 ise, bu ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131n ve bunlara verilen yan\u0131tlar\u0131n ger\u00e7ek zamanl\u0131 veya a\u015famal\u0131 olarak i\u015flenmesini ifade eder. Geleneksel olarak, bir model t\u00fcm yan\u0131t\u0131n\u0131 olu\u015fturur ve sonra tek seferde sunard\u0131. Ancak ak\u0131\u015fl\u0131 sistemlerde, model yan\u0131t\u0131n\u0131n par\u00e7alar\u0131n\u0131 (token token) \u00fcretirken ayn\u0131 zamanda ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 da belirtebilir ve bu \u00e7a\u011fr\u0131lar e\u015f zamanl\u0131 olarak y\u00fcr\u00fct\u00fclebilir. Bu yakla\u015f\u0131m, \u00f6zellikle kullan\u0131c\u0131 deneyimi a\u00e7\u0131s\u0131ndan b\u00fcy\u00fck avantajlar sunar. Kullan\u0131c\u0131, modelin d\u00fc\u015f\u00fcnme veya i\u015flem yapma s\u00fcrecini anl\u0131k olarak takip edebilir, yan\u0131t\u0131n tamamlanmas\u0131n\u0131 beklemek zorunda kalmaz. \u00d6rne\u011fin, bir chatbotun “Hava durumu bilgisini al\u0131yorum…” demesi ve ayn\u0131 anda arka planda hava durumu API’sini \u00e7a\u011f\u0131rmas\u0131, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131s\u0131na bir \u00f6rnektir. Bu s\u00fcre\u00e7, kullan\u0131c\u0131ya s\u00fcrekli geri bildirim sa\u011flayarak etkile\u015fimi daha ak\u0131c\u0131 ve do\u011fal hale getirir.<\/p>\n
Peki, bu neden bu kadar \u00f6nemlidir? \u00d6ncelikle, kullan\u0131c\u0131 deneyimini (user experience) devrim niteli\u011finde iyile\u015ftirir. Kullan\u0131c\u0131lar, uzun bekleme s\u00fcreleri yerine anl\u0131k ve kademeli yan\u0131tlar alarak sistemle daha derin bir ba\u011f kurar. \u0130kincil olarak, karma\u015f\u0131k i\u015f ak\u0131\u015flar\u0131n\u0131n (workflow) y\u00f6netilmesini kolayla\u015ft\u0131r\u0131r. Bir yapay zeka arac\u0131s\u0131 (agent), birden fazla arac\u0131 ard\u0131\u015f\u0131k veya paralel olarak \u00e7a\u011f\u0131rabilir ve her \u00e7a\u011fr\u0131n\u0131n sonucunu bir sonraki ad\u0131ma girdi olarak kullanabilir. Bu, \u00e7ok ad\u0131ml\u0131 otomasyon s\u00fcre\u00e7lerinin geli\u015ftirilmesine olanak tan\u0131r. \u00dc\u00e7\u00fcnc\u00fcs\u00fc, performans optimizasyonu sa\u011flar. Uzun s\u00fcren bir API \u00e7a\u011fr\u0131s\u0131 bile, di\u011fer i\u015flemlerin e\u015f zamanl\u0131 olarak y\u00fcr\u00fct\u00fclmesine engel olmaz, b\u00f6ylece genel yan\u0131t s\u00fcresi (latency) azal\u0131r. Ancak bu avantajlar, beraberinde karma\u015f\u0131k y\u00f6netim sorunlar\u0131n\u0131 da getirir. Hata i\u015fleme (error handling), durum y\u00f6netimi (state management), e\u015f zamanl\u0131l\u0131k (concurrency) ve g\u00f6zlemlenebilirlik (observability) gibi konular, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131n ba\u015far\u0131l\u0131 bir \u015fekilde uygulanmas\u0131 i\u00e7in kritik \u00f6neme sahiptir.<\/p>\n
Ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131n temelini, genellikle “ak\u0131ll\u0131 arac\u0131lar” (smart agents) ve “fonksiyon \u00e7a\u011fr\u0131lar\u0131” (function calls) olu\u015fturur. Ak\u0131ll\u0131 arac\u0131lar, bir B\u00fcy\u00fck Dil Modeli’nin (LLM) etraf\u0131na in\u015fa edilmi\u015f, belirli hedeflere ula\u015fmak i\u00e7in \u00e7e\u015fitli ara\u00e7lar\u0131 kullanabilen otonom sistemlerdir. Bu arac\u0131lar, kullan\u0131c\u0131dan gelen girdiyi analiz eder, hangi ara\u00e7lar\u0131n (API’ler, veritaban\u0131 sorgular\u0131, dahili fonksiyonlar vb.) kullan\u0131lmas\u0131 gerekti\u011fine karar verir, bu ara\u00e7lar\u0131 \u00e7a\u011f\u0131r\u0131r, sonu\u00e7lar\u0131 i\u015fler ve kullan\u0131c\u0131ya anlaml\u0131 bir yan\u0131t sunar. Bu s\u00fcre\u00e7, genellikle bir “d\u00fc\u015f\u00fcnme-eylem-g\u00f6zlem” (think-act-observe) d\u00f6ng\u00fcs\u00fc i\u00e7inde ger\u00e7ekle\u015fir. LLM, bir d\u00fc\u015f\u00fcnce (thought) \u00fcretir, ard\u0131ndan bir eylem (action) olarak bir ara\u00e7 \u00e7a\u011fr\u0131s\u0131 yapar, bu eylemin sonucunu (observation) al\u0131r ve bir sonraki d\u00fc\u015f\u00fcncesini buna g\u00f6re \u015fekillendirir.<\/p>\n
Fonksiyon \u00e7a\u011fr\u0131lar\u0131 ise, LLM’lerin belirli bir formatta (genellikle JSON) ara\u00e7 \u00e7a\u011f\u0131rma niyetlerini ifade etme yetene\u011fidir. LLM’ye, kullanabilece\u011fi ara\u00e7lar\u0131n bir listesi ve her bir arac\u0131n ne i\u015fe yarad\u0131\u011f\u0131, hangi parametreleri ald\u0131\u011f\u0131 gibi bilgiler verilir. Model, bir kullan\u0131c\u0131 iste\u011fi ald\u0131\u011f\u0131nda, bu ara\u00e7lar\u0131 kullanarak en uygun yan\u0131t\u0131 olu\u015fturabilece\u011fini d\u00fc\u015f\u00fcn\u00fcrse, bir fonksiyon \u00e7a\u011fr\u0131s\u0131 tan\u0131m\u0131 \u00fcretir. Bu tan\u0131m, \u00e7a\u011fr\u0131lacak fonksiyonun ad\u0131n\u0131 ve gerekli parametreleri i\u00e7erir. \u00d6rne\u011fin, kullan\u0131c\u0131 “\u0130stanbul’da hava nas\u0131l?” diye sordu\u011funda, LLM, kendisine \u00f6\u011fretilen Ak\u0131\u015f ba\u011flam\u0131nda, bu s\u00fcre\u00e7 daha dinamik hale gelir. LLM, yan\u0131t\u0131n\u0131 par\u00e7a par\u00e7a \u00fcretirken, ayn\u0131 zamanda fonksiyon \u00e7a\u011fr\u0131s\u0131 tan\u0131m\u0131n\u0131 da par\u00e7a par\u00e7a yay\u0131nlayabilir. Geli\u015ftirici olarak biz, bu ak\u0131\u015f\u0131 dinleriz, fonksiyon \u00e7a\u011fr\u0131s\u0131 tan\u0131m\u0131n\u0131n tamamland\u0131\u011f\u0131n\u0131 alg\u0131lad\u0131\u011f\u0131m\u0131zda arac\u0131 y\u00fcr\u00fct\u00fcr\u00fcz ve arac\u0131n sonucunu tekrar LLM’ye veya do\u011frudan kullan\u0131c\u0131ya yay\u0131nlar\u0131z. Bu, karma\u015f\u0131k bir orkestrasyon (orchestration) gerektirir. Modelin \u00e7\u0131kt\u0131s\u0131n\u0131 s\u00fcrekli olarak izlemeli, ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 tan\u0131mal\u0131, bunlar\u0131 asenkron olarak y\u00fcr\u00fctmeli ve sonu\u00e7lar\u0131 do\u011fru yere y\u00f6nlendirmeliyiz. Bu, \u00f6zellikle birden fazla arac\u0131n paralel olarak \u00e7a\u011fr\u0131lmas\u0131 veya bir arac\u0131n sonucunun ba\u015fka bir arac\u0131 tetiklemesi gerekti\u011finde zorlay\u0131c\u0131 hale gelir. \u00d6rne\u011fin, bir kullan\u0131c\u0131 “Yar\u0131n \u0130stanbul’da hava nas\u0131l olacak ve bana bu bilgiyi e-posta ile g\u00f6nder” dedi\u011finde, arac\u0131 \u00f6nce hava durumu API’sini \u00e7a\u011f\u0131racak, ard\u0131ndan gelen bilgiyi kullanarak bir e-posta g\u00f6nderme arac\u0131n\u0131 \u00e7a\u011f\u0131racak ve bu s\u00fcre\u00e7lerin her biri ak\u0131\u015f halinde ger\u00e7ekle\u015febilir. Bu t\u00fcr \u00e7ok ad\u0131ml\u0131 ve e\u015f zamanl\u0131 senaryolar, sa\u011flam bir mimari ve dikkatli bir hata y\u00f6netimi gerektirir.<\/p>\n Ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 etkin bir \u015fekilde y\u00f6netmek, sisteminizin kararl\u0131l\u0131\u011f\u0131 ve performans\u0131 i\u00e7in hayati \u00f6neme sahiptir. Bu s\u00fcre\u00e7te kar\u015f\u0131la\u015fabilece\u011finiz zorluklar\u0131n \u00fcstesinden gelmek i\u00e7in belirli stratejiler ve tasar\u0131m kal\u0131plar\u0131 (design patterns) benimsemek gerekir. Temelde, bu stratejiler performans\u0131 art\u0131rma, hatalar\u0131 y\u00f6netme ve sistemin genel dayan\u0131kl\u0131l\u0131\u011f\u0131n\u0131 sa\u011flama \u00fczerine odaklan\u0131r.<\/p>\n Evet, kesinlikle m\u00fcmk\u00fcn ve hatta \u00e7o\u011fu zaman gereklidir. Ak\u0131\u015fl\u0131 sistemlerde, bir arac\u0131n yan\u0131t\u0131n\u0131 beklerken di\u011fer i\u015flemlerin duraklamas\u0131 kabul edilemez. Bu nedenle, asenkron programlama (asynchronous programming) ve paralel \u00e7a\u011fr\u0131lar, performans\u0131 art\u0131rman\u0131n anahtar\u0131d\u0131r. Asenkron programlama, \u00f6zellikle Python’da Bir yapay zeka arac\u0131s\u0131, LLM’den gelen ak\u0131\u015fta birden fazla ba\u011f\u0131ms\u0131z ara\u00e7 \u00e7a\u011fr\u0131s\u0131 tespit etti\u011finde, bunlar\u0131 s\u0131rayla (sequential) de\u011fil, paralel olarak (parallel) y\u00fcr\u00fctmelidir. \u00d6rne\u011fin, bir kullan\u0131c\u0131 hem hava durumu hem de en son haberler hakk\u0131nda bilgi istedi\u011finde, bu iki API \u00e7a\u011fr\u0131s\u0131 e\u015f zamanl\u0131 olarak ba\u015flat\u0131labilir. Bu, toplam yan\u0131t s\u00fcresini \u00f6nemli \u00f6l\u00e7\u00fcde azalt\u0131r. Ancak paralel \u00e7a\u011fr\u0131lar, dikkatli bir kaynak y\u00f6netimi gerektirir. \u00c7ok fazla paralel \u00e7a\u011fr\u0131, sistem kaynaklar\u0131n\u0131 (bellek, CPU, a\u011f ba\u011flant\u0131lar\u0131) t\u00fcketebilir ve harici API’ler i\u00e7in h\u0131z limitlerine (rate limits) tak\u0131lman\u0131za neden olabilir. Bu nedenle, bir havuz (pool) mekanizmas\u0131 kullanarak paralel \u00e7a\u011fr\u0131 say\u0131s\u0131n\u0131 s\u0131n\u0131rlamak ve ak\u0131ll\u0131ca \u00f6nceliklendirme yapmak \u00f6nemlidir. Ayr\u0131ca, her bir paralel \u00e7a\u011fr\u0131n\u0131n ba\u015far\u0131s\u0131z olma olas\u0131l\u0131\u011f\u0131n\u0131 da g\u00f6z \u00f6n\u00fcnde bulundurarak, hata i\u015fleme mekanizmalar\u0131n\u0131n bu senaryoya uygun oldu\u011fundan emin olmal\u0131s\u0131n\u0131z. \u00d6rne\u011fin, bir hava durumu API’si ba\u015far\u0131s\u0131z olsa bile, haberler API’sinin sonucunu kullan\u0131c\u0131ya g\u00f6sterebilmelisiniz.<\/p>\n Yukar\u0131daki \u00f6rnekte, Ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131nda durum y\u00f6netimi (state management) ve hata tolerans\u0131 (fault tolerance) kritik \u00f6neme sahiptir. Bir i\u015flem birden fazla ad\u0131mda ger\u00e7ekle\u015fti\u011finde veya birden fazla arac\u0131 i\u00e7erdi\u011finde, her ad\u0131m\u0131n ve her \u00e7a\u011fr\u0131n\u0131n durumunu takip etmek zorunlu hale gelir. Hangi ara\u00e7 \u00e7a\u011fr\u0131s\u0131 ba\u015flat\u0131ld\u0131? Hangisi tamamland\u0131? Hangisi ba\u015far\u0131s\u0131z oldu? Bu sorular\u0131n yan\u0131tlar\u0131n\u0131 bilmek, sistemin do\u011fru \u00e7al\u0131\u015fmas\u0131n\u0131 ve kurtar\u0131labilir olmas\u0131n\u0131 sa\u011flar.<\/p>\n Durum y\u00f6netimi i\u00e7in, her ara\u00e7 \u00e7a\u011fr\u0131s\u0131na benzersiz bir kimlik (ID) atayabilir ve bu ID ile \u00e7a\u011fr\u0131n\u0131n durumunu (\u00f6rne\u011fin, Hata tolerans\u0131, sistemin beklenmeyen durumlara (API kesintileri, a\u011f hatalar\u0131, yanl\u0131\u015f parametreler) ra\u011fmen \u00e7al\u0131\u015fmaya devam edebilme yetene\u011fidir. Bu ba\u011flamda, a\u015fa\u011f\u0131daki mekanizmalar \u00f6nemlidir:<\/p>\n Bu stratejilerin birle\u015fimi, ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 y\u00f6neten sistemlerinizi daha sa\u011flam, g\u00fcvenilir ve s\u00fcrd\u00fcr\u00fclebilir hale getirecektir.<\/p>\n \u015eimdiye kadar teorik kavramlar\u0131 ele ald\u0131k. Gelin, bu kavramlar\u0131 Python kullanarak basit bir ak\u0131\u015f agent’\u0131 (arac\u0131s\u0131) \u00f6rne\u011fiyle somutla\u015ft\u0131ral\u0131m. Bu \u00f6rnek, bir LLM’den (basitle\u015ftirilmi\u015f bir sim\u00fclasyon) gelen ak\u0131\u015f\u0131 nas\u0131l dinleyece\u011fimizi, ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 nas\u0131l tespit edip y\u00fcr\u00fctece\u011fimizi ve sonu\u00e7lar\u0131 nas\u0131l geri yay\u0131nlayaca\u011f\u0131m\u0131z\u0131 g\u00f6sterecek. Amac\u0131m\u0131z, ak\u0131\u015f\u0131n kesintisizli\u011fini korurken, ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131n dinamik olarak i\u015flenmesini sa\u011flamakt\u0131r.<\/p>\n \u00d6rne\u011fimizde, bir B\u00fcy\u00fck Dil Modeli’nin \u00e7\u0131kt\u0131s\u0131n\u0131 sim\u00fcle eden bir asenkron jenerat\u00f6r (async generator) kullanaca\u011f\u0131z. Bu jenerat\u00f6r, hem metin par\u00e7ac\u0131klar\u0131 hem de ara\u00e7 \u00e7a\u011fr\u0131s\u0131 tan\u0131mlar\u0131 (JSON format\u0131nda) yay\u0131nlayacak. Agent’\u0131m\u0131z, bu ak\u0131\u015f\u0131 okuyacak, bir ara\u00e7 \u00e7a\u011fr\u0131s\u0131 tespit etti\u011finde ilgili arac\u0131 \u00e7al\u0131\u015ft\u0131racak ve arac\u0131n sonucunu tekrar ak\u0131\u015fa dahil edecektir. Bu, kullan\u0131c\u0131ya hem LLM’in d\u00fc\u015f\u00fcncelerini hem de ara\u00e7lar\u0131n eylemlerini e\u015f zamanl\u0131 olarak sunma yetene\u011fi kazand\u0131r\u0131r.<\/p>\n \u0130lk olarak, kullanaca\u011f\u0131m\u0131z ara\u00e7lar\u0131 tan\u0131mlayal\u0131m. Bu \u00f6rnekte, basit bir hava durumu alma arac\u0131 ve bir e-posta g\u00f6nderme arac\u0131 (her ikisi de sim\u00fcle edilmi\u015f) kullanaca\u011f\u0131z. Her ara\u00e7, asenkron bir fonksiyon olacak ve belirli parametreler alacakt\u0131r.<\/p>\n Yukar\u0131daki kod \u00f6rne\u011fi, temel bir ak\u0131\u015f agent’\u0131n\u0131n nas\u0131l \u00e7al\u0131\u015fabilece\u011fini g\u00f6stermektedir. Bu \u00f6rnek, ger\u00e7ek d\u00fcnya senaryolar\u0131n\u0131n karma\u015f\u0131kl\u0131\u011f\u0131n\u0131 basitle\u015ftirse de, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131s\u0131 y\u00f6netiminin temel mimarisini ve i\u015fleyi\u015fini anlamak i\u00e7in iyi bir ba\u015flang\u0131\u00e7 noktas\u0131d\u0131r. Ger\u00e7ek bir uygulamada, LLM entegrasyonu, daha karma\u015f\u0131k JSON ayr\u0131\u015ft\u0131rma, hata i\u015fleme ve durum y\u00f6netimi mekanizmalar\u0131 gibi ek bile\u015fenlere ihtiya\u00e7 duyulacakt\u0131r.<\/p>\n Basit bir ak\u0131\u015f agent’\u0131 olu\u015fturmak ba\u015flang\u0131\u00e7 i\u00e7in harika olsa da, \u00fcretim ortam\u0131nda (production environment) \u00e7al\u0131\u015fan, y\u00fcksek performansl\u0131 ve g\u00fcvenilir bir sistem kurmak i\u00e7in daha ileri d\u00fczey tekniklere ihtiya\u00e7 vard\u0131r. Bu b\u00f6l\u00fcmde, sisteminizin sa\u011fl\u0131\u011f\u0131n\u0131 izlemek ve performans\u0131n\u0131 s\u00fcrekli olarak iyile\u015ftirmek i\u00e7in kullanabilece\u011finiz stratejilere odaklanaca\u011f\u0131z.<\/p>\n Ak\u0131\u015f halindeki sistemler, do\u011fas\u0131 gere\u011fi dinamik ve da\u011f\u0131t\u0131k olduklar\u0131ndan, hata ay\u0131klamas\u0131 ve izlenmesi zor olabilir. Bir iste\u011fin sistemde hangi yollardan ge\u00e7ti\u011fini, hangi ara\u00e7lar\u0131n \u00e7a\u011fr\u0131ld\u0131\u011f\u0131n\u0131 ve bu \u00e7a\u011fr\u0131lar\u0131n ne kadar s\u00fcrd\u00fc\u011f\u00fcn\u00fc anlamak, sorun giderme ve performans analizi i\u00e7in kritik \u00f6neme sahiptir. \u0130\u015fte bu noktada g\u00f6zlemlenebilirlik (observability) devreye girer.<\/p>\n Sisteminiz b\u00fcy\u00fcd\u00fck\u00e7e ve daha fazla kullan\u0131c\u0131ya hizmet verdik\u00e7e, \u00f6l\u00e7eklenebilirlik (scalability) ve kaynak y\u00f6netimi (resource management) hayati \u00f6nem kazan\u0131r. Ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131, y\u00fcksek e\u015f zamanl\u0131l\u0131k ve I\/O yo\u011funlu\u011fu nedeniyle \u00f6zel dikkat gerektirir.<\/p>\n Bu ileri d\u00fczey teknikler, ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131s\u0131 y\u00f6netimini sadece \u00e7al\u0131\u015f\u0131r hale getirmekle kalmaz, ayn\u0131 zamanda onu \u00fcretim ortam\u0131n\u0131n zorluklar\u0131na kar\u015f\u0131 dayan\u0131kl\u0131, \u00f6l\u00e7eklenebilir ve y\u00f6netilebilir k\u0131lar.<\/p>\n Ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 y\u00f6netirken kar\u015f\u0131la\u015fabilece\u011finiz baz\u0131 ger\u00e7ek d\u00fcnya zorluklar\u0131 ve bu zorluklar\u0131n \u00fcstesinden gelmek i\u00e7in uygulanabilecek \u00e7\u00f6z\u00fcm yakla\u015f\u0131mlar\u0131n\u0131 vaka analizleri \u00fczerinden inceleyelim. Bu \u00f6rnekler, teorik bilgiyi pratik senaryolarla birle\u015ftirerek, kar\u015f\u0131la\u015fabilece\u011finiz sorunlara kar\u015f\u0131 sizi haz\u0131rlayacakt\u0131r.<\/p>\n Senaryo:<\/strong> Bir e-ticaret chatbotu, kullan\u0131c\u0131dan “En son indirimleri g\u00f6ster ve stokta ‘X’ \u00fcr\u00fcn\u00fc olup olmad\u0131\u011f\u0131n\u0131 kontrol et” gibi bir istek al\u0131r. Chatbotun hem bir indirim API’sine hem de bir stok y\u00f6netim API’sine ayn\u0131 anda \u00e7a\u011fr\u0131 yapmas\u0131 ve her iki API’den gelen yan\u0131tlar\u0131 h\u0131zl\u0131ca birle\u015ftirerek kullan\u0131c\u0131ya sunmas\u0131 gerekmektedir. Kullan\u0131c\u0131, chatbotun yan\u0131t\u0131n\u0131 beklerken s\u0131k\u0131lmamal\u0131 ve bilgi anl\u0131k olarak akmal\u0131d\u0131r.<\/p>\n Zorluklar:<\/strong><\/p>\n \u00c7\u00f6z\u00fcm Yakla\u015f\u0131m\u0131:<\/strong><\/p>\n Senaryo:<\/strong> Bir m\u00fc\u015fteri hizmetleri otomasyonu, karma\u015f\u0131k bir i\u015f ak\u0131\u015f\u0131n\u0131 y\u00f6netmektedir: m\u00fc\u015fteriden sorununu al\u0131r, dahili bir bilgi taban\u0131nda arama yapar, gerekirse bir destek bileti olu\u015fturur ve ard\u0131ndan m\u00fc\u015fteriye e-posta ile bilgilendirme yapar. Bu s\u00fcre\u00e7 birden fazla ara\u00e7 \u00e7a\u011fr\u0131s\u0131 i\u00e7erir ve toplamda birka\u00e7 dakika s\u00fcrebilir. E\u011fer bu s\u00fcre\u00e7 s\u0131ras\u0131nda sunucu \u00e7\u00f6ker veya a\u011f ba\u011flant\u0131s\u0131 kesilirse, s\u00fcrecin kald\u0131\u011f\u0131 yerden devam etmesi gerekmektedir.<\/p>\n Zorluklar:<\/strong><\/p>\n \u00c7\u00f6z\u00fcm Yakla\u015f\u0131m\u0131:<\/strong><\/p>\n Bu vaka analizleri, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131 d\u00fcnyas\u0131nda kar\u015f\u0131la\u015fabilece\u011finiz karma\u015f\u0131k sorunlara kar\u015f\u0131 pratik ve uygulanabilir \u00e7\u00f6z\u00fcm yakla\u015f\u0131mlar\u0131 sunmaktad\u0131r. Do\u011fru mimari se\u00e7imleri ve sa\u011flam m\u00fchendislik prensipleriyle, bu zorluklar\u0131n \u00fcstesinden gelmek m\u00fcmk\u00fcnd\u00fcr.<\/p>\n Ak\u0131\u015f halindeki ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 y\u00f6netmek, modern yapay zeka destekli uygulamalar\u0131n ve otomasyon sistemlerinin temel bir gereklili\u011fidir. Bu s\u00fcre\u00e7, kullan\u0131c\u0131 deneyimini zenginle\u015ftirirken, ayn\u0131 zamanda geli\u015ftiriciler i\u00e7in karma\u015f\u0131k teknik zorluklar\u0131 da beraberinde getirir. Ancak bu makalede ele ald\u0131\u011f\u0131m\u0131z stratejiler \u2013 asenkron i\u015fleme, paralel \u00e7a\u011fr\u0131lar, sa\u011flam durum y\u00f6netimi, kapsaml\u0131 hata tolerans\u0131, g\u00f6zlemlenebilirlik ve performans optimizasyonu \u2013 sayesinde, bu zorluklar\u0131n \u00fcstesinden gelmek ve ak\u0131l sa\u011fl\u0131\u011f\u0131n\u0131z\u0131 koruyarak ba\u015far\u0131l\u0131 sistemler in\u015fa etmek m\u00fcmk\u00fcnd\u00fcr. Unutmay\u0131n ki, iyi bir mimari tasar\u0131m ve s\u00fcrekli izleme, sistemlerinizin hem \u00f6l\u00e7eklenebilir hem de g\u00fcvenilir olmas\u0131n\u0131 sa\u011flayacakt\u0131r.<\/p>\n SSS 1: Ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 ak\u0131\u015f halinde kullanmak ne zaman avantajl\u0131d\u0131r?<\/strong><\/p>\n Cevap:<\/strong> Ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131, \u00f6zellikle kullan\u0131c\u0131ya ger\u00e7ek zamanl\u0131 geri bildirim sa\u011flamak istedi\u011finizde, yapay zeka ajanlar\u0131n\u0131n birden fazla ad\u0131m\u0131 veya arac\u0131 dinamik olarak kullanmas\u0131 gerekti\u011finde ve genel yan\u0131t s\u00fcresini (latency) minimize etmek istedi\u011finizde avantajl\u0131d\u0131r. Chatbotlar, sanal asistanlar ve karma\u015f\u0131k otomasyon i\u015f ak\u0131\u015flar\u0131 gibi senaryolarda kritik \u00f6neme sahiptir.<\/p>\n SSS 2: Hangi programlama dilleri bu t\u00fcr uygulamalar i\u00e7in daha uygundur?<\/strong><\/p>\n Cevap:<\/strong> Asenkron programlama yetenekleri g\u00fc\u00e7l\u00fc olan diller, bu t\u00fcr uygulamalar i\u00e7in idealdir. Python ( SSS 3: G\u00fcvenlik endi\u015feleri nas\u0131l ele al\u0131nmal\u0131d\u0131r?<\/strong><\/p>\n Cevap:<\/strong> G\u00fcvenlik, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131nda b\u00fcy\u00fck \u00f6nem ta\u015f\u0131r. Harici API’lere yap\u0131lan \u00e7a\u011fr\u0131lar i\u00e7in kimlik do\u011frulama (authentication) ve yetkilendirme (authorization) mekanizmalar\u0131n\u0131 (API anahtarlar\u0131, OAuth tokenlar\u0131) dikkatlice y\u00f6netin. Hassas bilgileri (API anahtarlar\u0131, \u015fifreler) asla kod i\u00e7inde saklamay\u0131n; bunun yerine g\u00fcvenli ortam de\u011fi\u015fkenleri veya gizli y\u00f6netim servisleri (\u00f6rne\u011fin AWS Secrets Manager) kullan\u0131n. Ayr\u0131ca, LLM’in \u00fcretebilece\u011fi ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 do\u011frulamak ve k\u00f6t\u00fc niyetli (malicious) veya yetkisiz \u00e7a\u011fr\u0131lar\u0131 engellemek i\u00e7in giri\u015f do\u011frulama (input validation) ve g\u00fcvenlik duvarlar\u0131 (firewalls) uygulay\u0131n.<\/p>\n SSS 4: Ara\u00e7 \u00e7a\u011fr\u0131lar\u0131nda veri tutarl\u0131l\u0131\u011f\u0131 nas\u0131l sa\u011flan\u0131r?<\/strong><\/p>\n Cevap:<\/strong> Veri tutarl\u0131l\u0131\u011f\u0131, \u00f6zellikle birden fazla arac\u0131 i\u00e7eren i\u015flemlerde zorlay\u0131c\u0131 olabilir. \u0130dempotans (idempotency) prensibini benimseyerek, ayn\u0131 \u00e7a\u011fr\u0131n\u0131n birden fazla kez yap\u0131lmas\u0131n\u0131n istenmeyen yan etkilere yol a\u00e7mamas\u0131n\u0131 sa\u011flay\u0131n. \u0130\u015flemleri m\u00fcmk\u00fcn oldu\u011funca atomik (atomic) hale getirin veya da\u011f\u0131t\u0131lm\u0131\u015f i\u015flemler (distributed transactions) i\u00e7in saga veya telafi edici i\u015flem (compensating transaction) desenlerini kullan\u0131n. Ayr\u0131ca, her arac\u0131n kendi veri tutarl\u0131l\u0131\u011f\u0131n\u0131 sa\u011flamas\u0131ndan emin olun ve olas\u0131 \u00e7ak\u0131\u015fmalar\u0131 veya tutars\u0131zl\u0131klar\u0131 ele alacak bir mekanizma geli\u015ftirin.<\/p>\n SSS 5: B\u00fcy\u00fck \u00f6l\u00e7ekli uygulamalarda performans darbo\u011fazlar\u0131 nas\u0131l a\u015f\u0131l\u0131r?<\/strong><\/p>\n Cevap:<\/strong> B\u00fcy\u00fck \u00f6l\u00e7ekli uygulamalarda performans darbo\u011fazlar\u0131n\u0131 a\u015fmak i\u00e7in kapsaml\u0131 \u00f6nbellekleme (caching), ak\u0131ll\u0131 h\u0131z s\u0131n\u0131rlama (rate limiting), y\u00fck dengeleme (load balancing) ve asenkron mesaj kuyruklar\u0131 (message queues) gibi teknikler kullan\u0131lmal\u0131d\u0131r. Mikroservis (microservices) mimarisi, farkl\u0131 ara\u00e7 \u00e7a\u011fr\u0131s\u0131 t\u00fcrlerini veya i\u015f ak\u0131\u015flar\u0131n\u0131 izole ederek ba\u011f\u0131ms\u0131z olarak \u00f6l\u00e7eklenmelerine olanak tan\u0131r. Ayr\u0131ca, s\u00fcrekli performans testi ve profilleme yaparak darbo\u011fazlar\u0131 erken tespit etmek ve optimize etmek \u00f6nemlidir.<\/p>\n #YapayZeka #APIY\u00f6netimi #Ak\u0131\u015fl\u0131Veri #AgentGeli\u015ftirme #Python #AsenkronProgramlama<\/p>\n","protected":false},"excerpt":{"rendered":"Modern yaz\u0131l\u0131m geli\u015ftirmenin ve \u00f6zellikle yapay zeka destekli sistemlerin kalbinde, farkl\u0131 servislerin ve ara\u00e7lar\u0131n birbiriyle etkile\u015fim kurmas\u0131 yatar.","protected":false},"author":1,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"csco_page_header_type":"","csco_page_load_nextpost":"","csco_page_subscribe_form":"","csco_page_contact_form":"","footnotes":""},"categories":[1],"tags":[],"class_list":{"0":"post-44187","1":"post","2":"type-post","3":"status-publish","4":"format-standard","6":"category-genel","7":"cs-entry","8":"cs-video-wrap"},"yoast_head":"\nget_weather(location: str)<\/code> fonksiyonunu \u00e7a\u011f\u0131raca\u011f\u0131n\u0131 ve location<\/code> parametresine “\u0130stanbul” de\u011ferini atayaca\u011f\u0131n\u0131 belirten bir JSON \u00e7\u0131kt\u0131s\u0131 \u00fcretebilir.<\/p>\nAk\u0131\u015f Halindeki Ara\u00e7 \u00c7a\u011fr\u0131lar\u0131n\u0131 Etkin Y\u00f6netme Stratejileri Nelerdir?<\/h2>\n
Asenkron \u0130\u015fleme ve Paralel \u00c7a\u011fr\u0131larla Performans\u0131 Art\u0131rmak M\u00fcmk\u00fcn m\u00fc?<\/h3>\n
asyncio<\/code> k\u00fct\u00fcphanesi ile pop\u00fclerle\u015fmi\u015ftir, bir fonksiyonun uzun s\u00fcreli bir i\u015flem (\u00f6rne\u011fin bir a\u011f iste\u011fi) yaparken CPU’yu bloke etmemesini sa\u011flar. Bu, ayn\u0131 anda birden fazla I\/O (giri\u015f\/\u00e7\u0131k\u0131\u015f) yo\u011funlu\u011fa sahip i\u015flemin y\u00fcr\u00fct\u00fclmesine olanak tan\u0131r.<\/p>\n\nimport asyncio\nimport httpx # Modern HTTP client for async operations\nimport json\n\nasync def fetch_url(url: str):\n async with httpx.AsyncClient() as client:\n response = await client.get(url, timeout=5)\n response.raise_for_status() # Raises an exception for 4xx\/5xx responses\n return response.json()\n\nasync def get_multiple_data(urls: list[str]):\n tasks = [fetch_url(url) for url in urls]\n results = await asyncio.gather(*tasks, return_exceptions=True) # Run tasks concurrently\n \n for i, result in enumerate(results):\n if isinstance(result, Exception):\n print(f\"HATA: {urls[i]} adresinden veri al\u0131namad\u0131: {result}\")\n else:\n print(f\"Ba\u015far\u0131l\u0131: {urls[i]} -> {json.dumps(result, indent=2)}\")\n\n# \u00d6rnek kullan\u0131m (ger\u00e7ek bir senaryoda bu URL'ler ara\u00e7 \u00e7a\u011fr\u0131lar\u0131ndan gelirdi)\n# async def main_example():\n# urls_to_fetch = [\n# \"https:\/\/api.weather.com\/v1\/current?location=istanbul\", # Mock URL\n# \"https:\/\/newsapi.org\/v2\/top-headlines?country=tr\" # Mock URL\n# ]\n# await get_multiple_data(urls_to_fetch)\n\n# if __name__ == \"__main__\":\n# asyncio.run(main_example())\n <\/code><\/pre>\n<\/div>\nasyncio.gather<\/code> kullanarak birden fazla HTTP iste\u011finin nas\u0131l paralel olarak yap\u0131labilece\u011fi g\u00f6sterilmi\u015ftir. Bu, ak\u0131\u015fl\u0131 ara\u00e7 \u00e7a\u011fr\u0131lar\u0131nda performans darbo\u011fazlar\u0131n\u0131 a\u015fmak i\u00e7in temel bir y\u00f6ntemdir.<\/p>\nDurum Y\u00f6netimi ve Hata Tolerans\u0131 Nas\u0131l Sa\u011flan\u0131r?<\/h3>\n
PENDING<\/code>, SUCCESS<\/code>, FAILED<\/code>) bir veritaban\u0131nda veya h\u0131zl\u0131 eri\u015fimli bir \u00f6nbellekte (Redis gibi) saklayabilirsiniz. Bu, sistemin \u00e7\u00f6kmesi veya yeniden ba\u015flat\u0131lmas\u0131 durumunda bile, devam eden i\u015flemleri kurtarman\u0131za olanak tan\u0131r. Ayr\u0131ca, bir kullan\u0131c\u0131 oturumu boyunca yap\u0131lan t\u00fcm ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131n ge\u00e7mi\u015fini tutmak, ba\u011flamsal (contextual) bir deneyim sunmak ve hata ay\u0131klamay\u0131 kolayla\u015ft\u0131rmak a\u00e7\u0131s\u0131ndan faydal\u0131d\u0131r.<\/p>\n\n
Pratik Uygulamalar: Python ile Basit Bir Ak\u0131\u015f Agent’\u0131 Olu\u015fturma<\/h2>\n
\nimport asyncio\nimport json\nimport time\n\n# --- Mock Ara\u00e7 Kay\u0131t Defteri (Tool Registry) ---\nasync def get_weather(location: str):\n \"\"\"Belirtilen konum i\u00e7in hava durumu bilgisini getirir.\"\"\"\n print(f\"DEBUG: 'get_weather' arac\u0131 \u00e7a\u011fr\u0131ld\u0131, konum: {location}\")\n await asyncio.sleep(1.5) # API \u00e7a\u011fr\u0131s\u0131 gecikmesini sim\u00fcle et\n if location.lower() == \"istanbul\":\n return {\"location\": \"\u0130stanbul\", \"temperature\": \"25\u00b0C\", \"condition\": \"G\u00fcne\u015fli\"}\n elif location.lower() == \"ankara\":\n return {\"location\": \"Ankara\", \"temperature\": \"20\u00b0C\", \"condition\": \"Par\u00e7al\u0131 Bulutlu\"}\n else:\n return {\"error\": \"Bilinmeyen konum veya hava durumu bilgisi mevcut de\u011fil.\"}\n\nasync def send_email(recipient: str, subject: str, body: str):\n \"\"\"Belirtilen al\u0131c\u0131ya e-posta g\u00f6nderir.\"\"\"\n print(f\"DEBUG: 'send_email' arac\u0131 \u00e7a\u011fr\u0131ld\u0131, al\u0131c\u0131: {recipient}, konu: {subject}\")\n await asyncio.sleep(2) # E-posta g\u00f6nderme gecikmesini sim\u00fcle et\n if \"@\" in recipient:\n return {\"status\": \"success\", \"message\": f\"'{subject}' konulu e-posta '{recipient}' adresine g\u00f6nderildi.\"}\n else:\n return {\"status\": \"error\", \"message\": \"Ge\u00e7ersiz e-posta adresi.\"}\n\ntool_registry = {\n \"get_weather\": get_weather,\n \"send_email\": send_email\n}\n\n# --- Mock LLM \u00c7\u0131kt\u0131 Ak\u0131\u015f\u0131 (Sim\u00fclasyon) ---\nasync def mock_llm_output_stream():\n \"\"\"LLM'den gelen ak\u0131\u015f\u0131 sim\u00fcle eder: metin ve ara\u00e7 \u00e7a\u011fr\u0131lar\u0131.\"\"\"\n yield \"Merhaba! \"\n await asyncio.sleep(0.1)\n yield \"Size nas\u0131l yard\u0131mc\u0131 olabilirim?\\n\"\n await asyncio.sleep(0.1)\n yield \"\u0130stanbul'daki hava durumu bilgisini \u00f6\u011frenmek istiyorsunuz, de\u011fil mi?\\n\"\n await asyncio.sleep(0.2)\n \n # LLM, get_weather arac\u0131n\u0131 \u00e7a\u011f\u0131r\u0131yor\n tool_call_json = {\n \"type\": \"tool_call\",\n \"name\": \"get_weather\",\n \"args\": {\"location\": \"Istanbul\"}\n }\n yield json.dumps(tool_call_json) + \"\\n\" # Ara\u00e7 \u00e7a\u011fr\u0131s\u0131 JSON'\u0131n\u0131 yay\u0131nla\n await asyncio.sleep(0.1)\n \n yield \"Hava durumu bilgisini al\u0131yorum... \"\n await asyncio.sleep(0.5)\n \n # LLM, e-posta g\u00f6nderme arac\u0131n\u0131 da \u00e7a\u011f\u0131rabilir (\u00f6rne\u011fin hava durumunu g\u00f6nder)\n # Bu \u00f6rnekte basitle\u015ftirmek i\u00e7in tek \u00e7a\u011fr\u0131 ile ilerleyelim.\n # Ger\u00e7ek senaryoda LLM, hava durumu sonucunu g\u00f6rd\u00fckten sonra e-posta \u00e7a\u011fr\u0131s\u0131 yapabilir.\n \n yield \"Ba\u015fka bir iste\u011finiz var m\u0131?\"\n await asyncio.sleep(0.1)\n yield \"\\n\"\n\n# --- Ak\u0131\u015f Agent'\u0131 Ana \u0130\u015fleyici (Main Handler) ---\nasync def streaming_agent_handler():\n \"\"\"LLM \u00e7\u0131kt\u0131s\u0131n\u0131 i\u015fler, ara\u00e7 \u00e7a\u011fr\u0131lar\u0131n\u0131 y\u00fcr\u00fct\u00fcr ve sonu\u00e7lar\u0131 yay\u0131nlar.\"\"\"\n current_tool_call_buffer = \"\"\n is_parsing_tool_call = False\n\n print(\"Agent ak\u0131\u015f\u0131 ba\u015flat\u0131l\u0131yor...\\n\")\n\n async for chunk in mock_llm_output_stream():\n # LLM'den gelen ham par\u00e7ay\u0131 do\u011frudan yay\u0131nla (kullan\u0131c\u0131ya anl\u0131k geri bildirim)\n yield chunk\n \n # Ara\u00e7 \u00e7a\u011fr\u0131s\u0131 JSON'\u0131n\u0131 tespit etme ve ayr\u0131\u015ft\u0131rma\n if '\"type\": \"tool_call\"' in chunk:\n is_parsing_tool_call = True\n current_tool_call_buffer += chunk\n elif is_parsing_tool_call and '\\n' in chunk: # JSON bir sat\u0131rda bitiyorsa\n try:\n tool_call_data = json.loads(current_tool_call_buffer.strip())\n if tool_call_data.get(\"type\") == \"tool_call\":\n tool_name = tool_call_data[\"name\"]\n tool_args = tool_call_data[\"args\"]\n\n if tool_name in tool_registry:\n print(f\"\\n[AGENT]: '{tool_name}' arac\u0131 y\u00fcr\u00fct\u00fcl\u00fcyor...\")\n # Arac\u0131 asenkron olarak \u00e7a\u011f\u0131r\n tool_result = await tool_registry[tool_name](**tool_args)\n \n # Ara\u00e7 sonucunu ak\u0131\u015fa dahil et\n yield json.dumps({\"type\": \"tool_result\", \"name\": tool_name, \"result\": tool_result}) + \"\\n\"\n print(f\"[AGENT]: '{tool_name}' arac\u0131 sonucu yay\u0131nland\u0131.\")\n else:\n yield json.dumps({\"type\": \"tool_error\", \"name\": tool_name, \"error\": \"Ara\u00e7 bulunamad\u0131\"}) + \"\\n\"\n print(f\"[AGENT]: HATA - '{tool_name}' arac\u0131 bulunamad\u0131.\")\n \n except json.JSONDecodeError:\n yield json.dumps({\"type\": \"tool_error\", \"error\": \"Ge\u00e7ersiz ara\u00e7 \u00e7a\u011fr\u0131s\u0131 JSON'\u0131\"}) + \"\\n\"\n print(f\"[AGENT]: HATA - Ge\u00e7ersiz ara\u00e7 \u00e7a\u011fr\u0131s\u0131 JSON'\u0131: {current_tool_call_buffer}\")\n finally:\n current_tool_call_buffer = \"\"\n is_parsing_tool_call = False\n elif is_parsing_tool_call: # JSON hen\u00fcz bitmediyse, tampona eklemeye devam et\n current_tool_call_buffer += chunk\n\n# --- Ana \u00c7al\u0131\u015ft\u0131rma Fonksiyonu ---\nasync def main():\n async for output_chunk in streaming_agent_handler():\n # Kullan\u0131c\u0131ya veya bir sonraki i\u015fleme \u00e7\u0131kt\u0131y\u0131 g\u00f6nder\n # Bu \u00f6rnekte do\u011frudan konsola yazd\u0131r\u0131yoruz\n print(output_chunk, end='')\n\n# if __name__ == \"__main__\":\n# asyncio.run(main())\n <\/code><\/pre>\n<\/div>\nmock_llm_output_stream<\/code> fonksiyonu, bir LLM’in hem metin hem de ara\u00e7 \u00e7a\u011fr\u0131lar\u0131 i\u00e7eren \u00e7\u0131kt\u0131s\u0131n\u0131 sim\u00fcle eder. tool_registry<\/code>, agent’\u0131m\u0131z\u0131n kullanabilece\u011fi ara\u00e7lar\u0131n bir haritas\u0131n\u0131 i\u00e7erir. streaming_agent_handler<\/code> ise, bu ak\u0131\u015f\u0131 i\u015fleyen ana mant\u0131kt\u0131r. Gelen her par\u00e7ay\u0131 analiz eder, bir ara\u00e7 \u00e7a\u011fr\u0131s\u0131 JSON’\u0131 tespit etti\u011finde onu ayr\u0131\u015ft\u0131r\u0131r, ilgili arac\u0131 asenkron olarak \u00e7a\u011f\u0131r\u0131r ve arac\u0131n sonucunu tekrar ak\u0131\u015fa JSON format\u0131nda dahil eder. Bu sayede, kullan\u0131c\u0131 hem LLM’in yan\u0131t\u0131n\u0131 hem de ara\u00e7lar\u0131n eylemlerini e\u015f zamanl\u0131 olarak deneyimler.<\/p>\n\u0130leri D\u00fczey Teknikler: G\u00f6zlemlenebilirlik ve Performans Optimizasyonu<\/h2>\n
Ak\u0131\u015f S\u00fcre\u00e7lerini \u0130zlemek ve Hata Ay\u0131klamak \u0130\u00e7in Neler Yapmal\u0131y\u0131z?<\/h3>\n
\n
\u00d6l\u00e7eklenebilirlik ve Kaynak Y\u00f6netimi Konusunda \u0130pu\u00e7lar\u0131<\/h3>\n
\n
Kar\u015f\u0131la\u015f\u0131lan Zorluklar ve \u00c7\u00f6z\u00fcm Yakla\u015f\u0131mlar\u0131: Vaka Analizleri<\/h2>\n
Vaka 1: Ger\u00e7ek Zamanl\u0131 Bir Chatbotun Birden Fazla API’ye E\u015f Zamanl\u0131 \u00c7a\u011fr\u0131 Yapmas\u0131 ve Yan\u0131tlar\u0131 Birle\u015ftirmesi<\/h3>\n
\n
\n
asyncio<\/code> k\u00fct\u00fcphanesi veya Node.js’deki Promise’ler gibi asenkron programlama paradigmalar\u0131 kullan\u0131larak her iki API \u00e7a\u011fr\u0131s\u0131 da e\u015f zamanl\u0131 olarak ba\u015flat\u0131l\u0131r. Bu, \u00e7a\u011fr\u0131lar\u0131n birbirini beklemesini engeller.\n\nasync def get_discounts():\n # ... API \u00e7a\u011fr\u0131s\u0131 sim\u00fclasyonu ...\n await asyncio.sleep(1)\n return {\"discounts\": [\"%20 indirimli ti\u015f\u00f6rtler\", \"\u00dccretsiz kargo\"]}\n\nasync def check_stock(product_name: str):\n # ... API \u00e7a\u011fr\u0131s\u0131 sim\u00fclasyonu ...\n await asyncio.sleep(1.2)\n return {\"product\": product_name, \"in_stock\": True, \"quantity\": 50}\n\nasync def handle_multi_api_request(product: str):\n discount_task = asyncio.create_task(get_discounts())\n stock_task = asyncio.create_task(check_stock(product))\n\n discounts = await discount_task\n stock_info = await stock_task\n\n # Yan\u0131tlar\u0131 birle\u015ftirme\n response_parts = []\n if discounts:\n response_parts.append(f\"En son indirimlerimiz: {', '.join(discounts['discounts'])}\")\n if stock_info:\n response_parts.append(f\"'{stock_info['product']}' \u00fcr\u00fcn\u00fc stokta: {'Evet' if stock_info['in_stock'] else 'Hay\u0131r'}.\")\n \n return \" \".join(response_parts)\n <\/code><\/pre>\n<\/p><\/div>\n<\/li>\ntry-except<\/code>) kullan\u0131l\u0131r. Bir \u00e7a\u011fr\u0131 ba\u015far\u0131s\u0131z olsa bile, di\u011ferinin ba\u015far\u0131l\u0131 sonucunu kullan\u0131c\u0131ya sunmak ve ba\u015far\u0131s\u0131z olan \u00e7a\u011fr\u0131 hakk\u0131nda bilgi vermek (\u00f6rn. “Stok bilgisi \u015fu an al\u0131nam\u0131yor.”) \u00f6nemlidir. Devre kesiciler, s\u00fcrekli ba\u015far\u0131s\u0131z olan API’leri ge\u00e7ici olarak devre d\u0131\u015f\u0131 b\u0131rakmak i\u00e7in kullan\u0131labilir.<\/li>\n<\/ul>\nVaka 2: Uzun S\u00fcreli Bir Otomasyon Ak\u0131\u015f\u0131nda Ara\u00e7 \u00c7a\u011fr\u0131lar\u0131n\u0131n Durumunu Koruma ve Kurtarma<\/h3>\n
\n
\n
STARTED<\/code>, COMPLETED<\/code>, FAILED<\/code>) bir veritaban\u0131nda (PostgreSQL, MongoDB) veya kal\u0131c\u0131 bir anahtar-de\u011fer deposunda (Redis) saklan\u0131r. Her ad\u0131m tamamland\u0131\u011f\u0131nda, durum g\u00fcncellenir.\n\n# Basit bir durum depolama \u00f6rne\u011fi (ger\u00e7ekte veritaban\u0131 veya Redis kullan\u0131lmal\u0131)\nworkflow_states = {}\n\nasync def execute_step(workflow_id: str, step_name: str, tool_func, *args, **kwargs):\n workflow_states[workflow_id] = {\"current_step\": step_name, \"status\": \"STARTED\"}\n print(f\"DEBUG: {workflow_id} - '{step_name}' ad\u0131m\u0131 ba\u015flat\u0131ld\u0131.\")\n try:\n result = await tool_func(*args, **kwargs)\n workflow_states[workflow_id][\"status\"] = \"COMPLETED\"\n print(f\"DEBUG: {workflow_id} - '{step_name}' ad\u0131m\u0131 tamamland\u0131.\")\n return result\n except Exception as e:\n workflow_states[workflow_id][\"status\"] = \"FAILED\"\n workflow_states[workflow_id][\"error\"] = str(e)\n print(f\"DEBUG: {workflow_id} - '{step_name}' ad\u0131m\u0131 HATA: {e}\")\n raise # Hatay\u0131 yeniden f\u0131rlat\n\n# ... Bu fonksiyon her ad\u0131mda \u00e7a\u011fr\u0131l\u0131r\n <\/code><\/pre>\n<\/p><\/div>\n<\/li>\nSonu\u00e7 ve S\u0131k\u00e7a Sorulan Sorular<\/h2>\n
S\u0131k\u00e7a Sorulan Sorular<\/h3>\n
asyncio<\/code> ile), Node.js (Promise’ler ve async\/await<\/code> ile), Go (goroutine’ler ile) ve Rust (async\/await<\/code> ile) gibi diller, e\u015f zamanl\u0131 ve paralel i\u015flemleri etkin bir \u015fekilde y\u00f6netme kabiliyetleri nedeniyle \u00f6ne \u00e7\u0131kar.<\/p>\n