<\/div>\n

Merhaba,<\/p>\n

Bu i\u00e7eri\u011fimizde, AI modellerini dura\u011fan e\u011fitim verilerinin \u00f6tesine ta\u015f\u0131yarak, g\u00fcncel verilere eri\u015fmesini sa\u011flayan RAG (Retrieval-Augmented Generation) davran\u0131\u015f\u0131n\u0131 mercek alt\u0131na alacak ve nedir, ne de\u011fildir tam teferruatl\u0131 bir de\u011ferlendirmede bulunuyor olaca\u011f\u0131z.<\/p>\n

RAG(Retrieval-Augmented Generation) Nedir?<\/h4>\n

RAG; bir LLM’in, gelen soruya cevap vermeden \u00f6nce, kendi ezberiyle cevap vermesi yerine, harici ve g\u00fcvenilir bir bilgi\/veri kayna\u011f\u0131ndan(PDF’ler, veri tabanlar\u0131, web sayfalar\u0131 vs.)<\/span> ilgili bilgileri bulup, bu bilgiler \u0131\u015f\u0131\u011f\u0131nda cevap \u00fcretmesi davran\u0131\u015f\u0131d\u0131r.<\/p>\n

RAG’\u0131n ak\u0131\u015f mant\u0131\u011f\u0131 olduk\u00e7a basittir:<\/p>\n

<\/p>\n

\n \ud83d\udd04 RAG Ak\u0131\u015f Mant\u0131\u011f\u0131
\n <\/h3>\n

Sorgu (Query)<\/strong>
\nKullan\u0131c\u0131 bir soru sorar.\n <\/div>\n
\u2b07\ufe0f<\/div>\n
Arama (Retrieval)<\/strong>
\nSistem, \u015firket i\u00e7i belgelerinizde veya g\u00fcncel verilerde bu soruyla ilgili par\u00e7alar\u0131 arar.\n <\/div>\n
\u2b07\ufe0f<\/div>\n
Zenginle\u015ftirme (Augmentation)<\/strong>
\nBulunan bu ham bilgiler, kullan\u0131c\u0131n\u0131n sorusuyla birle\u015ftirilir.\n <\/div>\n
\u2b07\ufe0f<\/div>\n
\u00dcretim (Generation)<\/strong>
\nLLM, \u00f6n\u00fcne koyulan bu spesifik belgelere bakarak do\u011fru ve g\u00fcncel bir cevap yazar.\n <\/div>\n<\/div>\n
RAG, Ne Kadar Kritik Bir Davran\u0131\u015ft\u0131r?<\/h4>\n
Kurumsal ve ciddi sistemlerde RAG yoksa, sistem yok diyebiliriz.<\/em><\/p><\/blockquote>\n
RAG, kurumsal yapay zeka projeleri i\u00e7in hayati (mission-critical)<\/em> \u00f6neme sahiptir. Bir model ne kadar geli\u015fmi\u015f olursa olsun (GPT-4, Claude 3.5 vb.), e\u011fitildi\u011fi tarih itibar\u0131yla bilgi seti donacakt\u0131r. Haliyle RAG ile bu donmu\u015f bilgi setine canl\u0131l\u0131k kazand\u0131r\u0131labilmektedir.<\/p>\n
Bu a\u00e7\u0131dan bak\u0131ld\u0131\u011f\u0131nda; kurumsal chatbot’larda, i\u00e7 dok\u00fcman sorgulamalarda, mevzuat \/ hukuk \/ prosed\u00fcr i\u00e7eriklerinin analizinde, teknik destek botlar\u0131nda ve agent tabanl\u0131 geli\u015fmi\u015f sistemlerde RAG, opsiyonel bir eklentiden ziyade, bilgi a\u00e7\u0131s\u0131ndan g\u00fcvenilirlik i\u00e7in zorunlu bir reflekstir.<\/p>\n
RAG’s\u0131z LLM, g\u00fczel konu\u015fan ama sorumluluk almayan biri gibi davran\u0131r.<\/em><\/p><\/blockquote>\n
RAG olmasa ne olurdu?<\/strong>
\nModel bilmedi\u011fi bir konuda -bilmiyorum-<\/em> demek yerine, hal\u00fcsinasyona u\u011frayarak mant\u0131kl\u0131 g\u00f6r\u00fcnen ama tamamen yanl\u0131\u015f bilgiler uydurabilir. RAG bu tarz emin olunmayan yerlerdeki verilen cevaplarda, -Bu bilgiyi nereden ald\u0131n?-<\/em> sorusuna g\u00f6sterebilece\u011fi referans\u0131 olaca\u011f\u0131 i\u00e7in tutarl\u0131l\u0131k sa\u011flanabilmektedir. <\/p>\n
Bunun d\u0131\u015f\u0131nda LLM, bilgi eskimesinden dolay\u0131 \u015firketinizin d\u00fcnk\u00fc sat\u0131\u015f verilerini veya yeni yay\u0131nlanan bir y\u00f6netmeli\u011fi asla bilemez. RAG ile kuruma \u00f6zel bilgi s\u00fcrece dahil edilip; LLM’e \u015firket haf\u0131zas\u0131 kazand\u0131r\u0131labilir, anl\u0131k ve g\u00fcncel bilgi ile konu\u015fmas\u0131 sa\u011flanabilir.<\/p>\n
T\u00fcm bunlar\u0131n d\u0131\u015f\u0131nda, t\u00fcm verilerinizi modele en ba\u015ftan \u00f6\u011fretmek (fine-tuning) hem \u00e7ok pahal\u0131 olacakt\u0131r hem de verinin g\u00fcncelli\u011fi de\u011fi\u015ftik\u00e7e kontrol\u00fc zorla\u015facak, verisel a\u00e7\u0131dan karanl\u0131k bir alan olu\u015facakt\u0131r. \u0130\u015fte bu tarz duruma kar\u015f\u0131nda RAG, g\u00fcvenlik riskini olabildi\u011fince minimize etmektedir.<\/p>\n
RAG’\u0131n \u00c7\u00f6zd\u00fc\u011f\u00fc Ger\u00e7ek Sorunlar Nelerdir?<\/h4>\n
Bunu a\u015fa\u011f\u0131daki gibi bir ka\u00e7 ba\u015fl\u0131k alt\u0131nda, sahada kar\u015f\u0131la\u015f\u0131lan sorunlar e\u015fli\u011finde \u00f6rneklendirerek toparlayabiliriz;<\/p>\n
\n
G\u00fcncellik<\/em><\/strong>
\nMisal olarak RAG; -Bug\u00fcn y\u00fcr\u00fcrl\u00fc\u011fe giren yeni vergi kanunu nedir?-<\/em> sorusuna yan\u0131t verebilmeyi sa\u011flar.\n<\/li>\n
\u00d6zel Veri Eri\u015fimi<\/em><\/strong>
\n\u015eirketinize \u00f6zel, gizli ve internette bulunmayan dok\u00fcmanlar \u00fczerinden soru-cevap yap\u0131lmas\u0131n\u0131 sa\u011flar.\n<\/li>\n
Kaynak G\u00f6sterme<\/em><\/strong>
\nRAG sistemleri, cevab\u0131 hangi dok\u00fcman\u0131n hangi sayfas\u0131ndan ald\u0131\u011f\u0131na kadar detayl\u0131 referans verebilmektedir. Bu da g\u00fcvenilirlik a\u00e7\u0131s\u0131ndan olduk\u00e7a \u00f6nemlidir.<\/p>\n
Ayr\u0131ca LLM’den her \u015feyi bilmesini beklemek gereksizdir. Nereden bakaca\u011f\u0131n\u0131 bilmesi bizler i\u00e7in daha tercih edilebilirdir.<\/li>\n
D\u00fc\u015f\u00fck Maliyet<\/em><\/strong>
\nModeli yeniden e\u011fitmek yerine, sadece bir veri taban\u0131n\u0131 g\u00fcncellemek binlerce kat daha ucuz ve az maliyetlidir.\n<\/li>\n<\/ul>\n
\u00dcretilen Netice LLM B\u00fcnyesine Al\u0131n\u0131r m\u0131?<\/h4>\n
Hay\u0131r. RAG ile \u00fcretilen netice LLM’in b\u00fcnyesine (yani parametrelerine) al\u0131nmaz. Ancak bu sistemin asla \u00f6\u011frenemeyece\u011fi anlam\u0131na da gelmiyor. \u015e\u00f6yle ki, \u00f6ncelikle -LLM b\u00fcnyesine al\u0131nmaktan-<\/em> kastedilen nedir bunu \u00f6ncelikle anlayal\u0131m, ard\u0131ndan sistemin \u00f6\u011frenmesi<\/em> kavram\u0131yla aras\u0131ndaki ince ayr\u0131m\u0131 ortaya koymaya \u00e7al\u0131\u015fal\u0131m.<\/p>\n
“LLM b\u00fcnyesine al\u0131nmak” ne demek?<\/strong>
\nBir bilginin LLM b\u00fcnyesine al\u0131nmas\u0131 demek; modelin a\u011f\u0131rl\u0131klar\u0131n\u0131n (weights) de\u011fi\u015fmesi, yani fine-tuning \/ training yap\u0131lmas\u0131 demektir. B\u00f6ylece model kapat\u0131l\u0131p a\u00e7\u0131ld\u0131\u011f\u0131nda bile o bilginin kal\u0131c\u0131 olarak hat\u0131rlanmas\u0131 sa\u011flanm\u0131\u015f olacakt\u0131r. Evet, RAG’da bunlar\u0131n hi\u00e7biri s\u00f6z konusu de\u011fildir.<\/p>\n
RAG’de; model \u00f6\u011frenmez, a\u011f\u0131rl\u0131klar de\u011fi\u015fmez, kal\u0131c\u0131 haf\u0131za olu\u015fmaz ve sonraki konu\u015fmada \u00fcretilen neticeler otomatik olarak hat\u0131rlanmaz! Sadece o anki cevap, d\u0131\u015f kayna\u011fa dayan\u0131r ve ge\u00e7ici bir ba\u011flam \u00fczerinden model, sanki \u00f6nceden bu bilgiyi biliyormu\u015f gibi cevap \u00fcretir.<\/p>\n
RAG, modele okuyarak cevap \u00fcretmeyi sa\u011flarken; Fine-tuning ise ezberleyerek cevap \u00fcretmeyi \u00f6\u011frenmesini sa\u011flar.<\/em><\/p><\/blockquote>\n
Peki \u00fcretilen sonu\u00e7lar tamamen \u00e7\u00f6pe mi gidiyor?<\/strong>
\nTabi ki de hay\u0131r! Sonu\u00e7ta \u00fcretilen veriler LLM’e olmasa da sisteme kazand\u0131r\u0131l\u0131yor. Ne mi yap\u0131l\u0131yor?<\/em> \u00dcretilen ve do\u011frulanan veri dok\u00fcmana d\u00f6n\u00fc\u015ft\u00fcr\u00fcl\u00fcyor, embedding al\u0131n\u0131yor ve vector store’a yaz\u0131l\u0131yor. Bir sonraki soruda ise e\u011fer ihtiya\u00e7 varsa tekrar \u00e7ekilebiliyor. Yani b\u00f6ylece model olmasa da sistem \u00f6\u011freniyor<\/u>! Bu anlay\u0131\u015f, konu ba\u011flam\u0131nda olduk\u00e7a kritiktir!<\/p>\n
Ha \u00fcretilen cevab\u0131 model g\u00f6r\u00fcyor mu?<\/em> diye sorarsan\u0131z… evet, g\u00f6r\u00fcyor! Ancak malumunuz; g\u00f6rmek, \u00f6\u011frenmek demek de\u011fildir! LLM, prompt i\u00e7erisindeki her \u015feyi ge\u00e7ici ba\u011flam olarak g\u00f6rd\u00fc\u011f\u00fc i\u00e7in context window kapan\u0131nca her \u015fey unutuluyor!<\/p>\n
Ayr\u0131ca \u015funa da dikkatinizi \u00e7ekmek isterim ki, RAG sonucu elde edilen veriler tam do\u011frulanmadan LLM’e yaz\u0131lmas\u0131 da pek do\u011fru olmayacakt\u0131r. Ne de olsa RAG \u00e7\u0131kt\u0131s\u0131; yanl\u0131\u015f olabilir, ba\u011flama \u00f6zel olabilir ya da zamana ba\u011fl\u0131 olabilir. Bu tarz olas\u0131 durumlara sahip verileri LLM’e g\u00f6mmek demek modeli kal\u0131c\u0131 olarak kirletmek demektir. E\u011fer ki eldeki veri sabit, evrensel ve de\u011fi\u015fmeyecek bilgiyse i\u015fte o taktirde LLM b\u00fcnyesine aktar\u0131lma kayg\u0131s\u0131 g\u00fcd\u00fclmelidir.<\/p>\n
Bu a\u00e7\u0131dan bak\u0131ld\u0131\u011f\u0131nda sistemlerimizi tasarlarken do\u011fru mimari bizlere g\u00f6z k\u0131rpmakta ve bilgiyi d\u0131\u015fardan alan LLM modellemesi kar\u015f\u0131m\u0131za \u00e7\u0131kmaktad\u0131r. Evet, sistemin akl\u0131 LLM’de de\u011fil AI Agent’ta olmal\u0131d\u0131r. LLM olduk\u00e7a saf kalmal\u0131 ve prompt’a kar\u015f\u0131l\u0131k veri \u00fcretimiyle ilgilenmelidir. AI Agent ise t\u00fcm s\u00fcrecin karar\u0131n\u0131 ve stratejisini y\u00fcr\u00fctmelidir. \u0130deal olan sistem b\u00f6yledir…<\/p>\n
RAG’\u0131n MCP \u0130le Alakas\u0131 Var m\u0131?<\/h4>\n
RAG ile MCP<\/a> ayn\u0131 \u015fey olmasa da akraba iki teknolojidir diyebiliriz. MCP, LLM’e d\u0131\u015f d\u00fcnyay\u0131 standart bir \u015fekilde a\u00e7an protokold\u00fcr. Yani MCP, LLM’in d\u0131\u015far\u0131dan nas\u0131l beslenece\u011finin protokolvari cevab\u0131d\u0131r.<\/p>\n
\u015e\u00f6yle d\u00fc\u015f\u00fcn\u00fclebilir… RAG’da, orkestrasyon katman\u0131 olan AI Agent; ne zaman, hangi bilgiyi alaca\u011f\u0131n\u0131 sorarken; MCP’de ise yine orkestrasyon katman\u0131 bu bilgiyi LLM’e nas\u0131l verece\u011fini sormaktad\u0131r. Yani RAG bir davran\u0131\u015f ve stratejiyken, MCP ise ta\u015f\u0131ma ve entegrasyon standard\u0131d\u0131r diyebiliriz.<\/p>\n
MCP olmadan RAG olur mu?<\/strong>
\nEvet, zaten bu g\u00fcne kadar yap\u0131lan RAG’lar\u0131n %90’\u0131 zaten MCP’sizdi. Ancak MCP ile yap\u0131lan RAG’larda bir standartla\u015fma s\u00f6z konusu olmakta ve her tool i\u00e7in ayr\u0131 prompt yazma derdi sona ermektedir. Ancak unutulmamas\u0131 gereken nokta \u015fudur ki, MCP RAG’\u0131 garanti etmemektedir.<\/p>\n
Microsoft Agent Framework ile RAG Yap\u0131labilir mi?<\/h4>\n
Evet, hem de modelden ba\u011f\u0131ms\u0131z bir \u015fekilde… \u00c7\u00fcnk\u00fc RAG; model de\u011fil, bir orkestrasyon problemidir.<\/p>\n
RAG, LLM’in de\u011fil, AI Agent’\u0131n sorumlulu\u011fun da bir davran\u0131\u015ft\u0131r.<\/em><\/p><\/blockquote>\n
\u0130ster Azure OpenAI \u00fczerinden GPT modellerini kullan\u0131n, ister lokalde \u00e7al\u0131\u015fan bir Llama modelini… fark etmeksizin herhangi bir LLM ile Microsoft Agent Framework<\/a> arac\u0131l\u0131\u011f\u0131yla bu modellere bir RAG Tool tan\u0131mlanabilmektedir.<\/p>\n
\u00c7al\u0131\u015fma mant\u0131\u011f\u0131 olduk\u00e7a basittir; AI Agent’a \u015fu yetene\u011fin verilmesi yeterlidir: -E\u011fer cevab\u0131 bilmiyorsan, \u015fu veri taban\u0131na git ve ara-<\/em> Bu y\u00f6nlendirmeden sonra AI Agent, gerekti\u011fi taktirde belirlenen PDF veya SQL kayna\u011f\u0131ndan veriyi \u00e7ekecek ve yan\u0131t\u0131 ona g\u00f6re \u015fekillendirecektir.<\/p>\n
RAG yapmak i\u00e7in modelin kendisini de\u011fi\u015ftirmenize gerek yoktur! Yaln\u0131zca modelin \u00f6n\u00fcne bir -veri getirme mekanizmas\u0131- eklenmesi yeterli olacakt\u0131r.<\/em><\/p><\/blockquote>\n
\u00d6rnek \u00c7al\u0131\u015fma<\/strong>
\nMicrosoft Agent Framework ile a\u015fa\u011f\u0131daki gibi bir Vector Search<\/a> \u00e7al\u0131\u015fmas\u0131n\u0131 ele alarak RAG davran\u0131\u015f\u0131n\u0131 \u00f6rneklendirebiliriz. <\/p>\n
<\/p>\n
Projede kullan\u0131lan k\u00fct\u00fcphaneler\n <\/div>\n
<\/p>\n
\n
Microsoft.Agents.AI<\/a><\/li>\n
Microsoft.Agents.AI.OpenAI<\/a><\/li>\n
Microsoft.Agents.AI.Hosting<\/a><\/li>\n
Microsoft.Data.SqlClient<\/a><\/li>\n<\/ul><\/div>\n<\/div>\n
\u0130lk olarak \u00e7al\u0131\u015fman\u0131n sade halini a\u015fa\u011f\u0131ya alal\u0131m:<\/p>\n
\n
\r\n public class EmbeddingService\r\n {\r\n \/\/Embedding i\u015flemleri i\u00e7in standart bir aray\u00fczd\u00fcr. \r\n \/\/string verilen input'u, Embedding<float> t\u00fcr\u00fcnden output etmektedir.\r\n private readonly IEmbeddingGenerator<string, Embedding<float>> _embeddingGenerator;\r\n public EmbeddingService(string key, string model = "text-embedding-3-small", string endpoint = "https:\/\/openrouter.ai\/api\/v1")\r\n {\r\n \/\/OpenAI i\u00e7in EmbeddingClient olu\u015fturuluyoruz.\r\n var embeddingClient = new EmbeddingClient(\r\n model: model,\r\n credential: new System.ClientModel.ApiKeyCredential(key: key),\r\n options: new OpenAIClientOptions\r\n {\r\n Endpoint = new Uri(endpoint)\r\n }\r\n );\r\n\r\n \/\/Ve bu EmbeddingClient'\u0131 IEmbeddingGenerator'a d\u00f6n\u00fc\u015ft\u00fcrerek provider'dan ba\u011f\u0131ms\u0131z kod yaz\u0131labilir hale geliyoruz.\r\n _embeddingGenerator = embeddingClient.AsIEmbeddingGenerator();\r\n }\r\n\r\n \/\/ Tek bir metni embedding eden metottur.\r\n public async Task<float[]> GenerateEmbeddingAsync(string text)\r\n {\r\n var result = await _embeddingGenerator.GenerateAsync(text);\r\n\r\n return result.Vector.ToArray();\r\n }\r\n\r\n \/\/Daha h\u0131zl\u0131 bir embedding i\u015flemi sa\u011flayan metottur.\r\n public async Task<ReadOnlyMemory<float>> GenerateVectorAsync(string text)\r\n {\r\n return await _embeddingGenerator.GenerateVectorAsync(text);\r\n }\r\n\r\n public async Task<List<float[]>> GenerateEmbeddingsAsync(IEnumerable<string> texts)\r\n {\r\n var results = await _embeddingGenerator.GenerateAsync(texts);\r\n return results\r\n .Select(embedding => embedding.Vector.ToArray())\r\n .ToList();\r\n }\r\n\r\n \/\/\u0130ki vekt\u00f6r aras\u0131ndaki cosine similarity'i hesaplar.\r\n \/\/De\u011fer -1 ile 1 aras\u0131nda olacakt\u0131r. 1 = tatamen benzer, 0 = ili\u015fkisiz, -1 = z\u0131t\r\n public double CalculateCosineSimilarity(float[] vectors1, float[] vectors2)\r\n {\r\n \/\/Vekt\u00f6r boyutlar\u0131 e\u015fit olmal\u0131\r\n if (vectors1.Length != vectors2.Length)\r\n throw new ArgumentException("Vekt\u00f6r boyutlar\u0131 e\u015fle\u015fmiyor!");\r\n\r\n \/\/Dot product : \u0130ki vekt\u00f6r\u00fcn nokta \u00e7arp\u0131m\u0131\r\n \/\/\u03a3(a[i] * b[i])\r\n double dotProduct = 0;\r\n\r\n \/\/Magnitude (b\u00fcy\u00fckl\u00fck) hesaplamalar\u0131\r\n \/\/ ||a|| = \u221a(\u03a3(a[i]\u00b2))\r\n double mag1 = 0, mag2 = 0;\r\n\r\n \/\/T\u00fcm boyutlar i\u00e7in hesaplama\r\n for (int i = 0; i < vectors1.Length; i++)\r\n {\r\n dotProduct += vectors1[i] * vectors2[i]; \/\/ a[i] * b[i]\r\n mag1 += vectors1[i] * vectors1[i]; \/\/ a[i]\u00b2\r\n mag2 += vectors2[i] * vectors2[i]; \/\/ b[i]\u00b2\r\n }\r\n\r\n \/\/Magnitude'lar\u0131n karak\u00f6k\u00fcn\u00fc alma\r\n mag1 = Math.Sqrt(mag1);\r\n mag2 = Math.Sqrt(mag2);\r\n\r\n \/\/S\u0131f\u0131r vekt\u00f6r kontrol\u00fc\r\n if (mag1 == 0 || mag2 == 0)\r\n return 0;\r\n\r\n \/\/Cosine similarity form\u00fcl\u00fc: (a \u00b7 b) \/ (||a|| * ||b||)\r\n return dotProduct \/ (mag1 * mag2);\r\n }\r\n }\r\n<\/pre>\n<\/div>\n
Burada ilk olarak vector search s\u00fcrecinde embedding sorumlulu\u011funu \u00fcstlenen s\u0131n\u0131f\u0131m\u0131z\u0131 g\u00f6r\u00fcyoruz. Bu s\u0131n\u0131f i\u00e7erisindeki IEmbeddingGenerator<\/code> aray\u00fcz\u00fc sayesinde OpenAI, Azure OpenAI, Ollama, HuggingFace vs. gibi provider’lardan ba\u011f\u0131ms\u0131z bir \u015fekilde embedding s\u00fcre\u00e7leri y\u00fcr\u00fct\u00fclebilmekte ve EmbeddingClient<\/code> s\u0131n\u0131f\u0131 sayesinde de ilgili provider’\u0131n embedding endpoint’i ile ileti\u015fime ge\u00e7ilebilmektedir. S\u00fcre\u00e7te text-embedding-3-small<\/code> modeli e\u015fli\u011finde embedding s\u00fcrecinin seyretti\u011fini, ayr\u0131ca dikkat ederseniz; GenerateAsync<\/code> metodu ile embedding i\u015flemlerinin y\u00fcr\u00fct\u00fcld\u00fc\u011f\u00fcn\u00fc, bir yandan da GenerateVectorAsync<\/code> metoduyla daha az maliyetli embedding s\u00fcrecinin de desteklendi\u011fini g\u00f6receksiniz.<\/p>\n
\u015eimdi de embedding edilmi\u015f verilerin, veritaban\u0131 i\u015flemlerinin ve gerekli vector search \u00e7al\u0131\u015fmalar\u0131n\u0131n sorumlulu\u011funu \u00fcstlenecek veritaban\u0131 s\u0131n\u0131f\u0131n\u0131 ele alal\u0131m:<\/p>\n
\n\r\n public class SqlVectorStoreService\r\n {\r\n private readonly string _connectionString;\r\n\r\n public SqlVectorStoreService(string connectionString = "Server=localhost,1433;Database=ExampleDB;User Id=sa;Password=Test123!;TrustServerCertificate=True;")\r\n {\r\n _connectionString = connectionString;\r\n }\r\n\r\n public async Task InitializeDatabaseAsync()\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n var createTableSql = $@"\r\n IF NOT EXISTS (SELECT * FROM sys.tables WHERE name = 'Embeddings')\r\n BEGIN\r\n CREATE TABLE Embeddings (\r\n Id INT IDENTITY(1,1) PRIMARY KEY,\r\n Text NVARCHAR(MAX) NOT NULL,\r\n EmbeddingVector VECTOR(1536) NOT NULL,\r\n Metadata NVARCHAR(MAX) NULL,\r\n CreatedAt DATETIME2 DEFAULT GETDATE()\r\n );\r\n END";\r\n\r\n using var command = new SqlCommand(createTableSql, connection);\r\n await command.ExecuteNonQueryAsync();\r\n }\r\n\r\n public async Task SaveEmbeddingAsync(string text, double[] embedding, string? metadata = null)\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n \/\/Double array'i byte array'e \u00e7eviriyoruz.\r\n var embeddingBytes = new byte[embedding.Length * sizeof(double)];\r\n\r\n Buffer.BlockCopy(embedding, 0, embeddingBytes, 0, embeddingBytes.Length);\r\n\r\n var insertSql = $@"\r\n INSERT INTO Embeddings (Text, EmbeddingVector, Metadata)\r\n VALUES (@Text, @Embedding, @Metadata)";\r\n\r\n using var command = new SqlCommand(insertSql, connection);\r\n\r\n command.Parameters.AddWithValue("@Text", text);\r\n command.Parameters.AddWithValue("@Embedding", embeddingBytes);\r\n command.Parameters.AddWithValue("@Metadata", metadata ?? (object)DBNull.Value);\r\n\r\n await command.ExecuteNonQueryAsync();\r\n }\r\n\r\n public async Task SaveEmbeddingsBatchAsync(List<(string text, float[] embedding, string metadata)> items)\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n using var transaction = connection.BeginTransaction();\r\n\r\n try\r\n {\r\n foreach (var (text, embedding, metadata) in items)\r\n {\r\n var embeddingBytes = new byte[embedding.Length * sizeof(float)];\r\n Buffer.BlockCopy(embedding, 0, embeddingBytes, 0, embeddingBytes.Length);\r\n\r\n var insertSql = $@"\r\n INSERT INTO Embeddings (Text, EmbeddingVector, Metadata)\r\n VALUES (@Text, @Embedding, @Metadata)";\r\n\r\n using var command = new SqlCommand(insertSql, connection, transaction);\r\n\r\n command.Parameters.AddWithValue("@Text", text);\r\n command.Parameters.AddWithValue("@Metadata", metadata ?? (object)DBNull.Value);\r\n\r\n var sqlVector = new SqlVector<float>(embedding);\r\n command.Parameters.Add(new SqlParameter\r\n {\r\n ParameterName = "@Embedding",\r\n Value = sqlVector,\r\n SqlDbType = System.Data.SqlDbType.Vector\r\n });\r\n\r\n await command.ExecuteNonQueryAsync();\r\n }\r\n\r\n await transaction.CommitAsync();\r\n }\r\n catch (Exception ex)\r\n {\r\n await transaction.RollbackAsync();\r\n throw new Exception($"Toplu kay\u0131t hatas\u0131 : {ex.Message}", ex);\r\n }\r\n }\r\n\r\n public async Task<List<SearchResult>> SearchAsync(float[] queryEmbedding, EmbeddingService embeddingService, int topK = 5, double minSimilarity = 0.0)\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n var selectSql = "SELECT Id, Text, EmbeddingVector, Metadata, CreatedAt FROM Embeddings";\r\n\r\n using var command = new SqlCommand(selectSql, connection);\r\n\r\n using var reader = await command.ExecuteReaderAsync();\r\n var results = new List<SearchResult>();\r\n\r\n while (await reader.ReadAsync())\r\n {\r\n var id = reader.GetInt32(0);\r\n var text = reader.GetString(1);\r\n var embeddingBytes = (byte[])reader.GetValue(2);\r\n\r\n var metadata = reader.IsDBNull(3) ? null : reader.GetString(3);\r\n var createdAt = reader.GetDateTime(4);\r\n\r\n var embedding = new float[embeddingBytes.Length \/ sizeof(float)];\r\n Buffer.BlockCopy(embeddingBytes, 0, embedding, 0, embeddingBytes.Length);\r\n\r\n var similarity = embeddingService.CalculateCosineSimilarity(queryEmbedding, embedding);\r\n if (similarity >= minSimilarity)\r\n {\r\n results.Add(new()\r\n {\r\n Id = id,\r\n Text = text,\r\n CreatedAt = createdAt,\r\n Metadata = metadata,\r\n Similarity = similarity\r\n });\r\n }\r\n }\r\n\r\n return results\r\n .OrderByDescending(r => r.Similarity)\r\n .Take(topK)\r\n .ToList();\r\n }\r\n\r\n public async Task ClearAllAsync()\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n var deleteSql = "TRUNCATE TABLE Embeddings";\r\n\r\n using var command = new SqlCommand(deleteSql, connection);\r\n await command.ExecuteNonQueryAsync();\r\n }\r\n\r\n public async Task<int> GetCountAsync()\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n var countSql = "SELECT COUNT(*) FROM Embeddings";\r\n using var command = new SqlCommand(countSql, connection);\r\n\r\n return (int)(await command.ExecuteScalarAsync());\r\n }\r\n\r\n public async Task<List<SearchResult>> VectorSearchAsync(float[] queryEmbedding, int topK = 5, double minSimilarity = 0.0)\r\n {\r\n using var connection = new SqlConnection(_connectionString);\r\n await connection.OpenAsync();\r\n\r\n var selectSql = @"\r\n SELECT TOP (@TopK)\r\n Id,\r\n Text,\r\n EmbeddingVector,\r\n Metadata,\r\n CreatedAt,\r\n -- Cosine similarity hesapla: 1 - cosine distance\r\n (1.0 - VECTOR_DISTANCE('cosine', EmbeddingVector, @QueryVector)) AS Similarity\r\n FROM Embeddings\r\n WHERE (1.0 - VECTOR_DISTANCE('cosine', EmbeddingVector, @QueryVector)) >= @MinSimilarity\r\n ORDER BY VECTOR_DISTANCE('cosine', EmbeddingVector, @QueryVector) ASC";\r\n\r\n using var command = new SqlCommand(selectSql, connection);\r\n\r\n command.Parameters.AddWithValue("@TopK", topK);\r\n command.Parameters.AddWithValue("@MinSimilarity", minSimilarity);\r\n\r\n var queryVector = new SqlVector<float>(queryEmbedding);\r\n command.Parameters.Add(new SqlParameter\r\n {\r\n ParameterName = "@QueryVector",\r\n Value = queryVector,\r\n SqlDbType = System.Data.SqlDbType.Vector\r\n });\r\n\r\n using var reader = await command.ExecuteReaderAsync();\r\n\r\n var results = new List<SearchResult>();\r\n while (await reader.ReadAsync())\r\n {\r\n var id = reader.GetInt32(0);\r\n var text = reader.GetString(1);\r\n \/\/ EmbeddingVector'u okumaya gerek yok (zaten similarity hesapland\u0131)\r\n var metadata = reader.IsDBNull(3) ? null : reader.GetString(3);\r\n var createdAt = reader.GetDateTime(4);\r\n var similarity = reader.GetDouble(5); \/\/ SQL Server'da hesaplanan similarity\r\n\r\n \/\/ Sonu\u00e7 listesine ekle\r\n results.Add(new SearchResult\r\n {\r\n Id = id,\r\n Text = text,\r\n Similarity = similarity,\r\n Metadata = metadata,\r\n CreatedAt = createdAt\r\n });\r\n }\r\n\r\n return results;\r\n }\r\n }\r\n<\/pre>\n<\/div>\nBurada klasik veritaban\u0131 i\u015flemlerinin yan\u0131nda esas olarak VectorSearchAsync<\/code> metodu i\u00e7erisinde vector search \u00e7al\u0131\u015fmas\u0131 ger\u00e7ekle\u015ftirilmekte ve SQL Server 2025 ile gelen VECTOR_DISTANCE<\/code><\/a> fonksiyonu sayesinde cosine similarity hesaplanarak gerekli sorgulama ger\u00e7ekle\u015ftirilmektedir.<\/p>\n Ve son olarak da bu \u00e7al\u0131\u015fman\u0131n precompute embeddings<\/em> ve semantic search<\/em> i\u015flemlerinin y\u00fcr\u00fct\u00fcld\u00fc\u011f\u00fc ana endpoint’lerini inceleyelim:<\/p>\n \n\r\napp.MapGet("\/precompute", async () =>\r\n{\r\n EmbeddingService embeddingService = new();\r\n var vectorStore = new SqlVectorStoreService();\r\n await vectorStore.InitializeDatabaseAsync();\r\n\r\n var documents = new List<string>\r\n {\r\n "Yapay zeka, insan zekas\u0131n\u0131 taklit eden bilgisayar sistemleridir.",\r\n "Machine learning, yapay zekan\u0131n \u00f6nemli bir alt dal\u0131d\u0131r.",\r\n "Derin \u00f6\u011frenme, sinir a\u011flar\u0131 kullanarak \u00e7al\u0131\u015f\u0131r ve karma\u015f\u0131k desenleri \u00f6\u011frenir.",\r\n "Python, veri bilimi ve makine \u00f6\u011frenmesi i\u00e7in en pop\u00fcler programlama dilidir.",\r\n "SQL Server, Microsoft'un g\u00fc\u00e7l\u00fc ili\u015fkisel veritaban\u0131 y\u00f6netim sistemidir.",\r\n "C# ile .NET Framework kullanarak g\u00fcvenli ve \u00f6l\u00e7eklenebilir uygulamalar geli\u015ftirebilirsiniz.",\r\n "Hugging Face, do\u011fal dil i\u015fleme modelleri i\u00e7in en b\u00fcy\u00fck platformdur.",\r\n "Embedding vekt\u00f6rleri, metinlerin anlam\u0131n\u0131 say\u0131sal olarak temsil eder.",\r\n "Cosine similarity, iki vekt\u00f6r aras\u0131ndaki a\u00e7\u0131y\u0131 \u00f6l\u00e7erek benzerlik hesaplar.",\r\n "Semantic search, anlam bazl\u0131 arama yaparak daha ilgili sonu\u00e7lar bulur.",\r\n "ONNX, makine \u00f6\u011frenmesi modellerini farkl\u0131 platformlarda \u00e7al\u0131\u015ft\u0131rmak i\u00e7in a\u00e7\u0131k formatt\u0131r.",\r\n "all-MiniLM-L6-v2, 384 boyutlu embedding vekt\u00f6rleri \u00fcreten hafif bir modeldir."\r\n };\r\n\r\n var embeddings = await embeddingService.GenerateEmbeddingsAsync(documents);\r\n var batch = documents\r\n .Zip(embeddings, (text, emb) => (text, emb, metadata: (string)null))\r\n .ToList();\r\n\r\n await vectorStore.SaveEmbeddingsBatchAsync(batch!);\r\n});\r\n\r\napp.MapGet("\/semantic-search\/{text?}", async (string? text) =>\r\n{\r\n EmbeddingService embeddingService = new();\r\n var vectorStore = new SqlVectorStoreService();\r\n if (text is not null)\r\n {\r\n var embedded = await embeddingService.GenerateEmbeddingAsync(text!);\r\n var results = await vectorStore.VectorSearchAsync(embedded);\r\n\r\n return TypedResults.Json(results.ToList());\r\n }\r\n return null;\r\n});\r\n<\/pre>\n<\/div>\nEvet… \u015eimdi bu en yal\u0131n haliyle ele ald\u0131\u011f\u0131m\u0131z \u00e7al\u0131\u015fma \u00fczerinden RAG davran\u0131\u015f\u0131n\u0131 sergilemeye ba\u015flayabiliriz.<\/p>\n Bunun i\u00e7in a\u015fa\u011f\u0131daki gibi iki adet endpoint geli\u015ftirmemiz yeterli olacakt\u0131r;<\/p>\n Basit yan\u0131t \u00fcretimi i\u00e7in:<\/em><\/p>\n\n\r\n\/\/ RAG sorgu endpoint'i - Basit yan\u0131t \u00fcretimi i\u00e7in\r\napp.MapGet("\/ask\/{question?}", async (string? question, EmbeddingService embeddingService, SqlVectorStoreService sqlVectorStoreService, [FromKeyedServices("ragAgent")] AIAgent ragAgent) =>\r\n{\r\n int topK = 3;\r\n double minSimilarity = 0.3;\r\n try\r\n {\r\n \/\/Kullan\u0131c\u0131n\u0131n sorusunu embedding'e \u00e7eviriyoruz.\r\n var queryEmbedding = await embeddingService.GenerateVectorAsync(question!);\r\n \/\/Vector store'dan en benzer i\u00e7erikleri getirtiyoruz.\r\n var searchResults = await sqlVectorStoreService.VectorSearchAsync(queryEmbedding.ToArray(), topK, minSimilarity);\r\n\r\n if (searchResults is [])\r\n return Results.Ok(new\r\n {\r\n question,\r\n answer = "\u00dczg\u00fcn\u00fcm, bu soruyla ilgili bilgi bulunamad\u0131.",\r\n sources = new int[] { }\r\n });\r\n\r\n StringBuilder context = new();\r\n context.AppendLine("\u0130lgili dok\u00fcmanlar :");\r\n context.AppendLine();\r\n\r\n for (int i = 0; i < searchResults.Count; i++)\r\n {\r\n context.AppendLine($"[Dok\u00fcman {i + 1}] (Benzerlik :{searchResults[i].Similarity:F3})");\r\n context.AppendLine(searchResults[i].Text);\r\n context.AppendLine();\r\n }\r\n\r\n var ragPrompt = $@"\r\n Context bilgileri : {context}\r\n Kullan\u0131c\u0131 sorusu : {question} \r\n\r\n L\u00fctfen yukar\u0131daki context bilgilerine dayanarak kullan\u0131c\u0131n\u0131n sorusunu yan\u0131tla.\r\n Ayr\u0131ca cevab\u0131nda hangi dok\u00fcmanlar\u0131 kulland\u0131\u011f\u0131n\u0131 belirt.";\r\n\r\n var response = await ragAgent.RunAsync(ragPrompt);\r\n return Results.Ok(new\r\n {\r\n question,\r\n answer = response.Text,\r\n sources = searchResults.Select(r => new\r\n {\r\n r.Text,\r\n similarity = Math.Round(r.Similarity, 3),\r\n id = r.Id\r\n }).ToList()\r\n });\r\n\r\n }\r\n catch (Exception ex)\r\n {\r\n return Results.BadRequest(new { error = ex.Message });\r\n }\r\n});\r\n<\/pre>\n<\/div>\nDetayl\u0131 yan\u0131t \u00fcretimi i\u00e7in:<\/em><\/p>\n\n\r\n\/\/ RAG sorgu endpoint'i - Detayl\u0131 yan\u0131t \u00fcretimi i\u00e7in\r\napp.MapPost("\/ask", async (RagQueryRequest request, EmbeddingService embeddingService, SqlVectorStoreService sqlVectorStoreService, [FromKeyedServices("ragAgent")] AIAgent ragAgent) =>\r\n{\r\n try\r\n {\r\n \/\/Kullan\u0131c\u0131n\u0131n sorusunu embedding'e \u00e7eviriyoruz.\r\n var queryEmbedding = await embeddingService.GenerateVectorAsync(request.Question!);\r\n \/\/Vector store'dan en benzer i\u00e7erikleri getirtiyoruz.\r\n var searchResults = await sqlVectorStoreService.VectorSearchAsync(queryEmbedding.ToArray(), request.TopK ?? 3, request.MinSimilarity ?? 0.3d);\r\n\r\n if (searchResults is [])\r\n return Results.Ok(new RagQueryResponse\r\n {\r\n Question = request.Question,\r\n Answer = "\u00dczg\u00fcn\u00fcm, bu soruyla ilgili bilgi bulunamad\u0131. L\u00fctfen sorunuzu farkl\u0131 \u015fekilde ifade etmeyi deneyiniz.",\r\n Sources = [],\r\n SourceCount = 0\r\n });\r\n\r\n StringBuilder context = new();\r\n context.AppendLine("\u0130lgili dok\u00fcmanlar :");\r\n context.AppendLine();\r\n\r\n for (int i = 0; i < searchResults.Count; i++)\r\n {\r\n context.AppendLine($"[Dok\u00fcman {i + 1}] (Benzerlik :{searchResults[i].Similarity:F3})");\r\n context.AppendLine(searchResults[i].Text);\r\n context.AppendLine();\r\n }\r\n\r\n var systemPrompt = request.CustomSystemPrompt ?? $@"\r\n Sen bir RAG asistan\u0131s\u0131n. G\u00f6revin:\r\n 1. Sa\u011flanan kaynak ve dok\u00fcmanlar\u0131 dikkatlice analiz et\r\n 2. Kullan\u0131c\u0131n\u0131n sorusuna YALNIZCA bu kaynaklara dayanarak cevap ver\r\n 3. Cevab\u0131nda hangi kayna\u011f\u0131 kulland\u0131\u011f\u0131n\u0131 [Kaynak-X] ya da [Dok\u00fcman-X] \u015feklinde belirt\r\n 4. E\u011fer kaynaklarda cevap yoksa, bunu a\u00e7\u0131k\u00e7a s\u00f6yle\r\n 5. Kendi bilgilerinle tahmin yapma, sadece verilen bilgileri kullan";\r\n\r\n var userPrompt = $@"\r\n Context bilgileri : {context}\r\n Kullan\u0131c\u0131 sorusu : {request.Question} \r\n\r\n L\u00fctfen yukar\u0131daki kaynak ve dok\u00fcmanlara dayanarak kullan\u0131c\u0131n\u0131n sorusunu yan\u0131tla.\r\n Ayr\u0131ca cevab\u0131nda hangi dok\u00fcmanlar\u0131 kulland\u0131\u011f\u0131n\u0131 belirt.";\r\n\r\n var messages = new List<OpenAI.Chat.ChatMessage>\r\n {\r\n OpenAI.Chat.ChatMessage.CreateSystemMessage(systemPrompt),\r\n OpenAI.Chat.ChatMessage.CreateUserMessage(userPrompt)\r\n };\r\n\r\n var chatResponse = await ragAgent.RunAsync(messages);\r\n var contentList = chatResponse.Content.OfType<ChatMessageContentPart>();\r\n\r\n var response = new RagQueryResponse\r\n {\r\n Question = request.Question,\r\n Answer = string.Join("", contentList.Select(item => item.Text)) ?? "Yan\u0131t olu\u015fturulmad\u0131!",\r\n Sources = searchResults.Select((r, index) => new SourceInfo\r\n {\r\n Id = r.Id,\r\n Text = r.Text,\r\n Similarity = r.Similarity,\r\n Rank = index + 1,\r\n Metadata = r.Metadata\r\n }).ToList(),\r\n SourceCount = searchResults.Count\r\n };\r\n\r\n return Results.Ok(response);\r\n }\r\n catch (Exception ex)\r\n {\r\n return Results.BadRequest(new { error = ex.Message, stackTrace = ex.StackTrace });\r\n }\r\n});\r\n<\/pre>\n<\/div>\nHer iki endpoint’e de g\u00f6z atarsan\u0131z e\u011fer kullan\u0131c\u0131dan gelen soru LLM’e g\u00f6nderilmeden \u00f6nce semantic search ile konuyla ilgili dok\u00fcmanlar taranmakta ve similarity’si y\u00fcksek olan ilk veriler elde edilerek prompt’a aktar\u0131lmaktad\u0131r. B\u00f6ylece LLM’in kendi e\u011fitim verisi d\u0131\u015f\u0131nda kalan bilgiler, anlamsal arama yoluyla harici veri kaynaklar\u0131ndan elde edilerek modele verilmekte ve b\u00f6ylece belirli bir \u00e7er\u00e7evede cevap \u00fcretmesi gerekti\u011fi s\u00f6ylemektedir…<\/p>\nBu minvalde detayl\u0131 yan\u0131t \u00fcretimi ger\u00e7ekle\u015ftiren endpoint’in test neticesini a\u015fa\u011f\u0131daki g\u00f6rselden inceleyebilir ve LLM’i RAG davran\u0131\u015f\u0131 neticesinde nas\u0131l \u00f6zelle\u015ftirebildi\u011fimizi de\u011ferlendirebilirsiniz:<\/a><\/p>\n \u0130\u015fte bu kadar \ud83d\ude42<\/p>\n Nihai olarak; \nRAG davran\u0131\u015f\u0131n\u0131n, b\u00fcy\u00fck dil modellerini “her \u015feyi ezberlemi\u015f bilge” olmaktan \u00e7\u0131kar\u0131p, “g\u00fcvenilir kaynaklara h\u0131zl\u0131ca bakabilen, g\u00fcncel ve sorumlu bir dan\u0131\u015fman” haline getiren olduk\u00e7a etkili ve ekonomik bir yakla\u015f\u0131m oldu\u011funun ve modern kurumsal yapay zeka mimarilerinin omurgas\u0131n\u0131n art\u0131k RAG’s\u0131z d\u00fc\u015f\u00fcn\u00fclemeyece\u011finin fark\u0131na vard\u0131\u011f\u0131m\u0131z\u0131 s\u00f6yleyebilir ve i\u00e7eri\u011fimizi burada noktayabiliriz…<\/p>\n \u0130lgilenenlerin faydalanmas\u0131 dile\u011fiyle… \nSonraki yaz\u0131lar\u0131mda g\u00f6r\u00fc\u015fmek \u00fczere… \n\u0130yi \u00e7al\u0131\u015fmalar…<\/p>\n Not : \u00d6rnek \u00e7al\u0131\u015fmaya a\u015fa\u011f\u0131daki GitHub adresinden eri\u015febilirsiniz. \nhttps:\/\/github.com\/gncyyldz\/Microsoft.Agent.Framework.RAG.Example<\/a><\/p>\n","protected":false},"excerpt":{"rendered":" Merhaba, Bu i\u00e7eri\u011fimizde, AI modellerini dura\u011fan e\u011fitim verilerinin \u00f6tesine ta\u015f\u0131yarak, g\u00fcncel verilere eri\u015fmesini sa\u011flayan RAG (Retrieval-Augmented Generation) davran\u0131\u015f\u0131n\u0131 mercek alt\u0131na alacak ve nedir, ne de\u011fildir tam teferruatl\u0131 bir de\u011ferlendirmede bulunuyor olaca\u011f\u0131z. RAG(Retrieval-Augmented Generation) Nedir?...<\/p>\n","protected":false},"author":1,"featured_media":28332,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[5206,5222,5220],"tags":[5601,5600,5291,5499,5305,5599,5304,5602,5394],"class_list":["post-28319","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-net","category-llm","category-yapay-zeka-ai","tag-embeddingclient","tag-iembeddinggenerator","tag-mcp","tag-microsoft-agent-framework","tag-rag","tag-rag-nedir","tag-retrieval-augmented-generation","tag-text-embedding-3-small","tag-vector-search"],"amp_enabled":true,"_links":{"self":[{"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/posts\/28319","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/types\/post"}],"author":[{"embeddable":true,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/comments?post=28319"}],"version-history":[{"count":10,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/posts\/28319\/revisions"}],"predecessor-version":[{"id":28333,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/posts\/28319\/revisions\/28333"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/media\/28332"}],"wp:attachment":[{"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/media?parent=28319"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/categories?post=28319"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/www.gencayyildiz.com\/blog\/wp-json\/wp\/v2\/tags?post=28319"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}