İLERİ

Vector Search & Semantic Search

ES 9.x ile native vector search: dense_vector field, kNN/ANN, semantic_text field ve ELSER modeli ile hybrid search.

Seviye: İleri+ — Bu bölüm production deneyimi gerektirir.

Karar Rehberi

Durum	Öneri	Örnek veya gerekçe
dense_vector + kNN	Uygun: Semantic similarity arama	RAG document retrieval
Hybrid (kNN + BM25)	Uygun: Hem anlam hem kelime eşleşmesi	E-ticaret ürün arama
semantic_text field	Uygun: Built-in embedding (kolay setup)	Quick semantic POC
ELSER	Uygun: Elastic hosted, sparse vector	English content search
External embeddings	Uygun: Custom domain model	Tıbbi/hukuki domain
Quantization (int8)	Uygun: Memory azaltma (>1M vector)	Large-scale production

REST API

# dense_vector mapping
curl -X PUT "http://localhost:9200/products-semantic" -H "Content-Type: application/json" -d'
{
  "mappings": {
    "properties": {
      "name": { "type": "text" },
      "description": { "type": "text" },
      "description_embedding": {
        "type": "dense_vector",
        "dims": 384,
        "index": true,
        "similarity": "cosine"
      },
      "semantic_description": {
        "type": "semantic_text",
        "inference_id": "my-elser-endpoint"
      }
    }
  }
}'

# kNN search
curl -X GET "http://localhost:9200/products-semantic/_search" -H "Content-Type: application/json" -d'
{
  "knn": {
    "field": "description_embedding",
    "query_vector": [0.1, 0.2, -0.3, ...],
    "k": 10,
    "num_candidates": 100
  },
  "_source": ["name", "description"]
}'

# Hybrid search (kNN + text)
curl -X GET "http://localhost:9200/products-semantic/_search" -H "Content-Type: application/json" -d'
{
  "query": {
    "match": { "description": "comfortable running shoes" }
  },
  "knn": {
    "field": "description_embedding",
    "query_vector": [0.1, 0.2, ...],
    "k": 10,
    "num_candidates": 50,
    "boost": 0.5
  },
  "size": 20
}'

# semantic_text (otomatik embedding - ES 9.x)
curl -X POST "http://localhost:9200/products-semantic/_doc" -H "Content-Type: application/json" -d'
{
  "name": "Ultra Comfortable Running Shoe",
  "semantic_description": "Lightweight breathable mesh upper with responsive foam cushioning for long distance running"
}'

.NET Client

public class SemanticSearchService
{
    private readonly ElasticsearchClient _client;
    private readonly IEmbeddingService _embeddings;

    public async Task<List<Product>> HybridSearchAsync(string query)
    {
        var embedding = await _embeddings.GenerateAsync(query);

        var response = await _client.SearchAsync<Product>(s => s
            .Index("products-semantic")
            .Query(q => q.Match(m => m
                .Field(f => f.Description)
                .Query(query)))
            .Knn(k => k
                .Field("description_embedding")
                .QueryVector(embedding)
                .k(10)
                .NumCandidates(100))
            .Size(20));

        return response.Documents.ToList();
    }
}

Örnek: RAG (Retrieval-Augmented Generation) pipeline'ında: kullanıcı sorusu → embedding model → ES kNN search → top-5 doküman → LLM'e context olarak gönder. Hybrid search (text + vector) sadece vector'dan %30 daha yüksek recall sağlar.