Bili\u015fsel Mimari: \u015eef Analojisiyle Anlamak<\/h3>\n\n\n\n
Bir agent\u2019\u0131n nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 anlaman\u0131n en etkili yollar\u0131ndan biri, s\u00fcreci insani bir metaforla a\u00e7\u0131klamakt\u0131r. \u00d6rne\u011fin; yo\u011fun bir mutfakta \u00e7al\u0131\u015fan bir \u015fefi d\u00fc\u015f\u00fcn\u00fcn. Bu \u015fefin amac\u0131, lezzetli yemekler haz\u0131rlamakt\u0131r. Ancak bu sadece tarif uygulamakla s\u0131n\u0131rl\u0131 de\u011fildir. \u015eef;<\/p>\n\n\n\n
- M\u00fc\u015fteri sipari\u015flerini ve mevcut malzemeleri analiz eder,<\/li>
- Hangi yeme\u011fi yapaca\u011f\u0131na karar verir,<\/li>
- Yeme\u011fi haz\u0131rlar,<\/li>
- Geri bildirim veya malzeme eksikliklerine g\u00f6re s\u00fcreci ayarlar.<\/li><\/ul>\n\n\n\n
Bu d\u00f6ng\u00fc, bilgi toplama \u2192 planlama \u2192 uygulama \u2192 ayarlama \u015feklindedir. \u015eefin bu s\u00fcrekli \u00f6\u011frenen, karar alan ve uygulayan yap\u0131s\u0131 asl\u0131nda bir bili\u015fsel mimari<\/strong> \u00f6rne\u011fidir.<\/p>\n\n\n\n
Agent\u2019lar da benzer \u015fekilde \u00e7al\u0131\u015f\u0131r: bilgi toplar, i\u00e7sel reasoning yapar, d\u0131\u015f d\u00fcnyayla etkile\u015fime girer ve elde etti\u011fi \u00e7\u0131kt\u0131lara g\u00f6re sonraki ad\u0131m\u0131n\u0131 \u015fekillendirir.<\/p>\n\n\n\n
Orkestrasyon Katman\u0131: Ak\u0131l Y\u00fcr\u00fctmenin Kalbi<\/h3>\n\n\n\n
![\"\"](\"\/wp-content\/uploads\/2025\/03\/image-13.png\")
<\/figure>\n\n\n\nAgent\u2019lar\u0131n bili\u015fsel mimarisinin merkezinde yer alan orchestration layer<\/strong>, memory, state, reasoning ve planning gibi unsurlar\u0131 y\u00f6netir. Bu katman, agent\u2019\u0131n g\u00f6revleri nas\u0131l ele alaca\u011f\u0131n\u0131 belirler ve prompt engineering<\/strong> tekniklerinden faydalan\u0131r. Dil modelleri i\u00e7in geli\u015ftirilen bir\u00e7ok framework, agent\u2019lar\u0131n daha etkili planlama yapmas\u0131na ve eyleme ge\u00e7mesine olanak tan\u0131r.<\/p>\n\n\n\n
En pop\u00fcler reasoning framework\u2019leri \u015funlard\u0131r:<\/p>\n\n\n\n
- ReAct (Reason + Act)<\/strong>: Dil modelinin bir kullan\u0131c\u0131 sorgusuna hem d\u00fc\u015f\u00fcnerek hem de eyleme ge\u00e7erek yan\u0131t vermesini sa\u011flar. SOTA (state-of-the-art) modelleri bile geride b\u0131rakan performanslar sergilemi\u015ftir.<\/li>
- Chain-of-Thought (CoT)<\/strong>: Karma\u015f\u0131k g\u00f6revleri, ara ad\u0131mlarla par\u00e7alayarak d\u00fc\u015f\u00fcnme yetene\u011fi sunar. Self-consistency, active-prompting gibi varyantlar\u0131 mevcuttur.<\/li>
- Tree-of-Thoughts (ToT)<\/strong>: CoT’un daha geli\u015fmi\u015f versiyonudur. Farkl\u0131 d\u00fc\u015f\u00fcnce dallar\u0131n\u0131 paralel olarak ke\u015ffederek stratejik g\u00f6revlerde kullan\u0131l\u0131r.<\/li><\/ul>\n\n\n\n
ReAct Framework\u2019\u00fc ile Agent D\u00f6ng\u00fcs\u00fc \u00d6rne\u011fi<\/h3>\n\n\n\n
ReAct framework kullanan bir agent\u2019\u0131n ad\u0131m ad\u0131m nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 inceleyelim:<\/p>\n\n\n\n
- Kullan\u0131c\u0131 bir sorgu g\u00f6nderir.<\/li>
- Agent, ReAct d\u00f6ng\u00fcs\u00fcn\u00fc ba\u015flat\u0131r.<\/li>
- Model \u015fu yap\u0131da ad\u0131mlar olu\u015fturur:
- Question<\/strong>: Kullan\u0131c\u0131dan gelen sorgu<\/li>
- Thought<\/strong>: Modelin bir sonraki ad\u0131m hakk\u0131nda d\u00fc\u015f\u00fcnd\u00fc\u011f\u00fc \u015fey<\/li>
- Action<\/strong>: Yap\u0131lmas\u0131 gereken eylem (\u00f6rne\u011fin: [Flights, Search, Code, None])<\/li>
- Action input<\/strong>: Eylemin alaca\u011f\u0131 parametreler<\/li>
- Observation<\/strong>: Eylemin sonucu<\/li>
- (Bu d\u00f6ng\u00fc gerekirse bir\u00e7ok kez tekrarlanabilir.)<\/li>
- Final Answer<\/strong>: Kullan\u0131c\u0131ya iletilecek nihai yan\u0131t<\/li><\/ul><\/li><\/ol>\n\n\n\n
Bu s\u00fcre\u00e7te model, varsay\u0131lan bilgiye dayal\u0131 bir tahminde bulunmak yerine (hallucination), \u00f6rne\u011fin bir “Flights” arac\u0131yla ger\u00e7ek zamanl\u0131 bilgi toplar ve daha do\u011fru, g\u00fcvenilir bir cevap \u00fcretir.<\/p>\n\n\n\n
Agent\u2019larda Cevap Kalitesi ve Tool Se\u00e7imi<\/h3>\n\n\n\n
Agent\u2019lar\u0131n \u00fcretti\u011fi yan\u0131tlar\u0131n kalitesi; do\u011fru ak\u0131l y\u00fcr\u00fctme stratejilerini kullanmalar\u0131na, uygun tool\u2019lar\u0131 se\u00e7ebilmelerine ve bu tool\u2019lar\u0131n nas\u0131l tan\u0131mland\u0131\u011f\u0131na do\u011frudan ba\u011fl\u0131d\u0131r. T\u0131pk\u0131 taze malzeme ve m\u00fc\u015fteri geri bildirimi ile en iyi yeme\u011fi haz\u0131rlayan bir \u015fef gibi, agent\u2019lar da do\u011fru reasoning ve g\u00fcvenilir bilgi ile optimum sonu\u00e7lar sunar.<\/p>\n\n\n\n
Tool\u2019lar: D\u0131\u015f D\u00fcnya ile Ba\u011flant\u0131 Noktas\u0131<\/h2>\n\n\n\n
Dil modelleri bilgi i\u015fleme konusunda g\u00fc\u00e7l\u00fc olsalar da, d\u0131\u015f d\u00fcnya ile do\u011frudan etkile\u015fim kurma yetenekleri s\u0131n\u0131rl\u0131d\u0131r. Modeller, e\u011fitim verilerinin \u00f6tesinde bilgiye eri\u015femez. Ancak fonksiyonlar, uzant\u0131lar, veri depolar\u0131 ve eklentiler<\/strong>, bu a\u00e7\u0131\u011f\u0131 kapatmak i\u00e7in tasarlanm\u0131\u015f ara\u00e7lard\u0131r.<\/p>\n\n\n\n
Bu ara\u00e7lar sayesinde agent\u2019lar:<\/p>\n\n\n\n
- Ak\u0131ll\u0131 ev ayarlar\u0131n\u0131 de\u011fi\u015ftirebilir,<\/li>
- Takvim g\u00fcncelleyebilir,<\/li>
- Veritaban\u0131ndan bilgi alabilir,<\/li>
- E-posta g\u00f6nderebilir.<\/li><\/ul>\n\n\n\n
B\u00f6ylece agent mimarisi, modelin kapasitesini a\u015farak daha y\u00fcksek do\u011fruluk, h\u0131z ve g\u00fcvenilirlikle g\u00f6revleri tamamlayabilir.<\/p>\n\n\n\n
Sonu\u00e7<\/h2>\n\n\n\n
Agent\u2019lar, sadece tahmin yapan modellerin \u00f6tesine ge\u00e7erek hedef odakl\u0131 \u00e7al\u0131\u015fan otonom sistemler haline geliyor. Bili\u015fsel mimariler, reasoning framework\u2019leri ve d\u0131\u015f d\u00fcnyayla ba\u011flant\u0131 kurabilen tool yap\u0131lar\u0131 sayesinde, daha dinamik, daha g\u00fcvenilir ve daha \u00fcretken \u00e7\u00f6z\u00fcmler sunabiliyorlar.<\/p>\n\n\n\n
Bu d\u00f6n\u00fc\u015f\u00fcm, gelecekte yapay zek\u00e2 sistemlerinin daha fazla alanda aktif rol \u00fcstlenmesinin \u00f6n\u00fcn\u00fc a\u00e7\u0131yor. Bir sonraki ad\u0131mda, bu agent\u2019lar\u0131n ger\u00e7ek zamanl\u0131 veri kaynaklar\u0131yla nas\u0131l entegre olduklar\u0131n\u0131 detayl\u0131ca inceleyece\u011fiz.<\/p>\n","protected":false},"excerpt":{"rendered":"
Yapay zek\u00e2 alan\u0131ndaki en \u00f6nemli d\u00f6n\u00fc\u015f\u00fcmlerden biri, modellerin sadece tahmin yapan sistemler olmaktan \u00e7\u0131karak, belirli hedeflere ula\u015fabilen otonom agent\u2019lara d\u00f6n\u00fc\u015fmesidir. Bu d\u00f6n\u00fc\u015f\u00fcm\u00fcn merkezinde, bili\u015fsel mimariler (cognitive architectures), ak\u0131l y\u00fcr\u00fctme framework\u2019leri ve tool entegrasyonlar\u0131 yer almaktad\u0131r. Bili\u015fsel Mimari: \u015eef Analojisiyle Anlamak Bir agent\u2019\u0131n nas\u0131l \u00e7al\u0131\u015ft\u0131\u011f\u0131n\u0131 anlaman\u0131n en etkili yollar\u0131ndan biri, s\u00fcreci insani bir metaforla a\u00e7\u0131klamakt\u0131r. \u00d6rne\u011fin;\u2026 Read More »<\/a><\/span><\/p>\n","protected":false},"author":1,"featured_media":348,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":[],"categories":[1],"tags":[154,155,151,150,133],"_links":{"self":[{"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/posts\/353"}],"collection":[{"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcomments&post=353"}],"version-history":[{"count":1,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/posts\/353\/revisions"}],"predecessor-version":[{"id":355,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/posts\/353\/revisions\/355"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=\/wp\/v2\/media\/348"}],"wp:attachment":[{"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=%2Fwp%2Fv2%2Fmedia&parent=353"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=%2Fwp%2Fv2%2Fcategories&post=353"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/blog.firatyasar.com\/index.php?rest_route=%2Fwp%2Fv2%2Ftags&post=353"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}
- Thought<\/strong>: Modelin bir sonraki ad\u0131m hakk\u0131nda d\u00fc\u015f\u00fcnd\u00fc\u011f\u00fc \u015fey<\/li>
- Question<\/strong>: Kullan\u0131c\u0131dan gelen sorgu<\/li>
- Chain-of-Thought (CoT)<\/strong>: Karma\u015f\u0131k g\u00f6revleri, ara ad\u0131mlarla par\u00e7alayarak d\u00fc\u015f\u00fcnme yetene\u011fi sunar. Self-consistency, active-prompting gibi varyantlar\u0131 mevcuttur.<\/li>
- ReAct (Reason + Act)<\/strong>: Dil modelinin bir kullan\u0131c\u0131 sorgusuna hem d\u00fc\u015f\u00fcnerek hem de eyleme ge\u00e7erek yan\u0131t vermesini sa\u011flar. SOTA (state-of-the-art) modelleri bile geride b\u0131rakan performanslar sergilemi\u015ftir.<\/li>