docs(semantic_search): recommend multilingual-e5-large for Turkish (#22)

Different embedding model families need different prompt prefixes —
Gemini wants "task: ... | query: ..." and "title: ... | text: ...",
e5 wants "query: ..." / "passage: ...", and using the wrong one
silently degrades retrieval quality. Add EMBEDDING_PROMPT_STYLE
(gemini/e5/raw) so the prefix matches the chosen model.

Defaults: gemini for OpenRouter (matches the existing default
google/gemini-embedding-001), e5 for the local provider (matches
the recommended multilingual-e5-large setup). Both override via
env var or constructor.

Update README and .env.example to recommend intfloat/multilingual-
e5-large served by HuggingFace Text Embeddings Inference (one
docker run) as the Turkish-optimized local setup, with a clear env
var reference table. Ollama and OpenRouter remain documented as
alternatives.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

.env.example

@@ -89,15 +89,30 @@ OPENROUTER_API_KEY=sk-or-v1-your_openrouter_api_key_here
 # OPENROUTER_EMBEDDING_MODEL=google/gemini-embedding-001
 # OPENROUTER_EMBEDDING_DIMENSION=3072
 
-# --- Option B: Local OpenAI-compatible server (Ollama / llama.cpp / vLLM) -----
+# --- Option B: Local OpenAI-compatible server (no API key required) ----------
-# Uncomment to use your own server instead of OpenRouter (no API key required).
+# Recommended for Turkish: intfloat/multilingual-e5-large served by HuggingFace
-# Defaults target Ollama with nomic-embed-text. For Turkish, bge-m3 (1024 dims)
+# Text Embeddings Inference (TEI). One-line setup:
-# tends to work better — pull it with: `ollama pull bge-m3`
+#
+#   docker run -p 8080:80 ghcr.io/huggingface/text-embeddings-inference:latest \
+#       --model-id intfloat/multilingual-e5-large
+#
+# Then uncomment the block below. Other model families work too — set
+# EMBEDDING_PROMPT_STYLE to match: e5 / gemini / raw.
+#
 # EMBEDDING_PROVIDER=local
-# LOCAL_EMBEDDING_BASE_URL=http://localhost:11434/v1
+# LOCAL_EMBEDDING_BASE_URL=http://localhost:8080/v1
-# LOCAL_EMBEDDING_MODEL=nomic-embed-text
+# LOCAL_EMBEDDING_MODEL=intfloat/multilingual-e5-large
-# LOCAL_EMBEDDING_DIMENSION=768
+# LOCAL_EMBEDDING_DIMENSION=1024
+# EMBEDDING_PROMPT_STYLE=e5
 # LOCAL_EMBEDDING_API_KEY=                          # most local servers ignore this
+#
+# Ollama fallback (if you prefer Ollama and don't need top Turkish quality):
+#   ollama serve && ollama pull nomic-embed-text
+#   EMBEDDING_PROVIDER=local
+#   LOCAL_EMBEDDING_BASE_URL=http://localhost:11434/v1
+#   LOCAL_EMBEDDING_MODEL=nomic-embed-text
+#   LOCAL_EMBEDDING_DIMENSION=768
+#   EMBEDDING_PROMPT_STYLE=raw   # nomic uses its own search_query/search_document
 
 # =============================================================================
 # USAGE INSTRUCTIONS

README.md

@@ -181,20 +181,40 @@ Yargı MCP'yi Gemini CLI ile kullanmak için:
 
 ---
 <details>
-<summary>🧠 <strong>Semantik Arama (Opsiyonel - OpenRouter API)</strong></summary>
+<summary>🧠 <strong>Semantik Arama (Opsiyonel)</strong></summary>
 
-Yargı MCP, **semantik arama** özelliği ile kararları anlamsal olarak sıralayabilir. Bu özellik opsiyoneldir ve `OPENROUTER_API_KEY` ayarlandığında otomatik olarak etkinleşir.
+Yargı MCP, **semantik arama** özelliği ile kararları anlamsal olarak sıralayabilir. Opsiyoneldir; iki yoldan biri yapılandırıldığında otomatik etkinleşir:
+
+- **Yerel** (önerilen, ücretsiz): kendi makinenizdeki OpenAI-uyumlu embedding sunucusu (HuggingFace TEI, llama.cpp, Ollama, vLLM, LM Studio…)
+- **Hosted**: OpenRouter API anahtarı
 
 ### Semantik Arama Nasıl Çalışır?
 1. `initial_keyword` ile Bedesten API'den 100 karar çekilir
 2. `query` ile bu kararlar embedding modeli kullanılarak anlamsal olarak sıralanır
 3. En alakalı kararlar döndürülür
 
-### OpenRouter API Anahtarı Alma
+### Önerilen Türkçe Kurulumu (Yerel — `multilingual-e5-large`)
-1. [OpenRouter](https://openrouter.ai/) sitesine gidin
-2. Hesap oluşturun ve API anahtarı alın (ücretsiz kredi ile başlayabilirsiniz)
 
-### Claude Desktop için Yapılandırma
+`intfloat/multilingual-e5-large` Türkçe için kıyas ettiğimiz açık kaynak modeller arasında en iyilerinden. HuggingFace'in **Text Embeddings Inference (TEI)** sunucusuyla tek komutta ayağa kalkar ve OpenAI-uyumlu API sunar:
+
+```bash
+docker run -p 8080:80 ghcr.io/huggingface/text-embeddings-inference:latest \
+    --model-id intfloat/multilingual-e5-large
+```
+
+Sonra Yargı MCP'ye şu env vars'ları geçirin:
+
+```bash
+EMBEDDING_PROVIDER=local
+LOCAL_EMBEDDING_BASE_URL=http://localhost:8080/v1
+LOCAL_EMBEDDING_MODEL=intfloat/multilingual-e5-large
+LOCAL_EMBEDDING_DIMENSION=1024
+EMBEDDING_PROMPT_STYLE=e5
+```
+
+> ⚠️ **Önemli:** `EMBEDDING_PROMPT_STYLE=e5` şart — e5 modelleri `query:` / `passage:` öneki bekleyecek şekilde eğitilmiştir; yanlış önek sessizce kaliteyi düşürür.
+
+#### Claude Desktop örneği (yerel TEI)
 ```json
 {
   "mcpServers": {
@@ -202,35 +222,60 @@ Yargı MCP, **semantik arama** özelliği ile kararları anlamsal olarak sırala
       "command": "uvx",
       "args": ["yargi-mcp"],
       "env": {
-        "OPENROUTER_API_KEY": "sk-or-v1-xxx..."
+        "EMBEDDING_PROVIDER": "local",
+        "LOCAL_EMBEDDING_BASE_URL": "http://localhost:8080/v1",
+        "LOCAL_EMBEDDING_MODEL": "intfloat/multilingual-e5-large",
+        "LOCAL_EMBEDDING_DIMENSION": "1024",
+        "EMBEDDING_PROMPT_STYLE": "e5"
       }
     }
   }
 }
 ```
 
-### 5ire için Yapılandırma
+### Alternatif 1: Ollama (yerel, daha hafif kurulum)
-Tool ayarlarında **Environment Variables** alanına ekleyin:
+
+```bash
+ollama serve
+ollama pull nomic-embed-text   # 768 dim, İngilizce ağırlıklı
 ```
+
+```bash
+EMBEDDING_PROVIDER=local
+LOCAL_EMBEDDING_BASE_URL=http://localhost:11434/v1
+LOCAL_EMBEDDING_MODEL=nomic-embed-text
+LOCAL_EMBEDDING_DIMENSION=768
+EMBEDDING_PROMPT_STYLE=raw
+```
+
+> Ollama kütüphanesinde `multilingual-e5-large` doğrudan yok; Türkçe için TEI yolu daha doğru sonuç verir.
+
+### Alternatif 2: OpenRouter (hosted)
+
+```bash
 OPENROUTER_API_KEY=sk-or-v1-xxx...
+# İsteğe bağlı — varsayılan google/gemini-embedding-001 (3072 dim, ÜCRETLİ)
+# OPENROUTER_EMBEDDING_MODEL=...
+# OPENROUTER_EMBEDDING_DIMENSION=...
+# EMBEDDING_PROMPT_STYLE=gemini   # varsayılan
 ```
 
-### Gemini CLI için Yapılandırma
+API anahtarınızı [openrouter.ai/keys](https://openrouter.ai/keys) adresinden alın. Varsayılan model `google/gemini-embedding-001` artık ücretli — ücretsiz bir model seçerseniz `OPENROUTER_EMBEDDING_MODEL`, `OPENROUTER_EMBEDDING_DIMENSION` ve uygun `EMBEDDING_PROMPT_STYLE` değerlerini birlikte ayarlayın.
-```json
-{
-  "mcpServers": {
-    "yargi_mcp": {
-      "command": "uvx",
-      "args": ["yargi-mcp"],
-      "env": {
-        "OPENROUTER_API_KEY": "sk-or-v1-xxx..."
-      }
-    }
-  }
-}
-```
 
-> 💡 **Not:** `OPENROUTER_API_KEY` ayarlanmazsa semantik arama aracı görünmez, diğer 24 araç normal şekilde çalışmaya devam eder.
+### Yapılandırma Referansı
+
+| Env Var | Açıklama | Örnek |
+|---|---|---|
+| `EMBEDDING_PROVIDER` | `local` ise yerel sunucu, boş ise OpenRouter | `local` |
+| `EMBEDDING_PROMPT_STYLE` | `gemini` / `e5` / `raw` — modelin beklediği önek | `e5` |
+| `LOCAL_EMBEDDING_BASE_URL` | Yerel sunucunun OpenAI-uyumlu URL'i | `http://localhost:8080/v1` |
+| `LOCAL_EMBEDDING_MODEL` | Model adı | `intfloat/multilingual-e5-large` |
+| `LOCAL_EMBEDDING_DIMENSION` | Modelin çıktı boyutu (mutlaka eşleşmeli) | `1024` |
+| `OPENROUTER_API_KEY` | OpenRouter anahtarı (sadece hosted için) | `sk-or-v1-…` |
+| `OPENROUTER_EMBEDDING_MODEL` | OpenRouter model id'si | `google/gemini-embedding-001` |
+| `OPENROUTER_EMBEDDING_DIMENSION` | OpenRouter modelinin çıktı boyutu | `3072` |
+
+> 💡 **Not:** Hiçbir embedding sağlayıcı yapılandırılmazsa semantik arama aracı görünmez, diğer 24 araç normal şekilde çalışır.
 
 </details>
 

semantic_search/embedder.py

@@ -14,13 +14,50 @@ DEFAULT_DIMENSION = 3072
 
 # Local provider defaults — Ollama with nomic-embed-text out of the box.
 # Override via LOCAL_EMBEDDING_BASE_URL / LOCAL_EMBEDDING_MODEL /
-# LOCAL_EMBEDDING_DIMENSION when using a different server or model
+# LOCAL_EMBEDDING_DIMENSION when using a different server or model.
-# (e.g. llama.cpp's server, vLLM, LM Studio, or a different Ollama model
+# For Turkish, intfloat/multilingual-e5-large (1024 dims, prompt_style=e5)
-# such as bge-m3 — better for Turkish — at 1024 dimensions).
+# served via HuggingFace TEI is the recommended setup — see README.
 LOCAL_DEFAULT_BASE_URL = "http://localhost:11434/v1"
 LOCAL_DEFAULT_MODEL = "nomic-embed-text"
 LOCAL_DEFAULT_DIMENSION = 768
 
+# Prompt-template styles. Embedding models are trained with specific
+# prefixes — using the wrong style silently degrades retrieval quality.
+#   - "gemini": "task: {task} | query: {text}" / "title: {title} | text: {text}"
+#               (matches google/gemini-embedding-001, the OpenRouter default)
+#   - "e5":     "query: {text}" / "passage: {text}"
+#               (matches intfloat/multilingual-e5-* models — best for Turkish)
+#   - "raw":    no prefix; pass text through as-is
+PROMPT_STYLES = ("gemini", "e5", "raw")
+DEFAULT_PROMPT_STYLE = "gemini"
+
+
+def _format_query(prompt_style: str, query: str, task: str) -> str:
+    if prompt_style == "e5":
+        return f"query: {query}"
+    if prompt_style == "raw":
+        return query
+    # gemini (default)
+    return f"task: {task} | query: {query}"
+
+
+def _format_document(prompt_style: str, doc: str, title: str) -> str:
+    if prompt_style == "e5":
+        return f"passage: {doc}"
+    if prompt_style == "raw":
+        return doc
+    # gemini (default)
+    return f"title: {title} | text: {doc}"
+
+
+def _resolve_prompt_style(explicit: Optional[str], default: str) -> str:
+    style = (explicit or os.getenv("EMBEDDING_PROMPT_STYLE") or default).strip().lower()
+    if style not in PROMPT_STYLES:
+        raise ValueError(
+            f"Unknown EMBEDDING_PROMPT_STYLE {style!r}; expected one of {PROMPT_STYLES}"
+        )
+    return style
+
 
 def is_openrouter_available() -> bool:
     """Check if OpenRouter API key is available."""
@@ -66,19 +103,21 @@ class _BaseOpenAICompatibleEmbedder:
     client = None
     model: str = ""
     dimension: int = 0
+    prompt_style: str = DEFAULT_PROMPT_STYLE
 
     def encode_query(self, query: str, task: str = "search result") -> np.ndarray:
         """
-        Encode a search query.
+        Encode a search query. Prefix is selected by ``self.prompt_style``.
 
         Args:
             query: The search query text
-            task: Task type for prompt template
+            task: Task hint used by the gemini-style prefix; ignored for
+                e5/raw styles.
 
         Returns:
             Numpy array of embeddings (``self.dimension`` elements).
         """
-        text = f"task: {task} | query: {query}"
+        text = _format_query(self.prompt_style, query, task)
 
         try:
             response = self.client.embeddings.create(
@@ -119,7 +158,7 @@ class _BaseOpenAICompatibleEmbedder:
         texts = []
         for i, doc in enumerate(documents):
             title = titles[i] if titles and i < len(titles) else "none"
-            texts.append(f"title: {title} | text: {doc}")
+            texts.append(_format_document(self.prompt_style, doc, title))
 
         try:
             response = self.client.embeddings.create(
@@ -186,7 +225,12 @@ class OpenRouterEmbedder(_BaseOpenAICompatibleEmbedder):
         "X-Title": "Yargi MCP Server",
     }
 
-    def __init__(self, model: Optional[str] = None, dimension: Optional[int] = None):
+    def __init__(
+        self,
+        model: Optional[str] = None,
+        dimension: Optional[int] = None,
+        prompt_style: Optional[str] = None,
+    ):
         api_key = os.getenv("OPENROUTER_API_KEY")
         if not api_key:
             raise ValueError("OPENROUTER_API_KEY environment variable is not set")
@@ -206,10 +250,14 @@ class OpenRouterEmbedder(_BaseOpenAICompatibleEmbedder):
             "OPENROUTER_EMBEDDING_DIMENSION",
             DEFAULT_DIMENSION,
         )
+        # Default to gemini-style prefix for OpenRouter — matches the default
+        # google/gemini-embedding-001 model. Override via constructor or
+        # EMBEDDING_PROMPT_STYLE env var when picking a different model.
+        self.prompt_style = _resolve_prompt_style(prompt_style, "gemini")
 
         logger.info(
             f"OpenRouter Embedder initialized with model: {self.model} "
-            f"(dimension={self.dimension})"
+            f"(dimension={self.dimension}, prompt_style={self.prompt_style})"
         )
 
 
@@ -240,6 +288,7 @@ class LocalEmbedder(_BaseOpenAICompatibleEmbedder):
         model: Optional[str] = None,
         dimension: Optional[int] = None,
         api_key: Optional[str] = None,
+        prompt_style: Optional[str] = None,
     ):
         try:
             from openai import OpenAI
@@ -266,10 +315,15 @@ class LocalEmbedder(_BaseOpenAICompatibleEmbedder):
             "LOCAL_EMBEDDING_DIMENSION",
             LOCAL_DEFAULT_DIMENSION,
         )
+        # Default to e5 prefix for local — the recommended Turkish setup
+        # (multilingual-e5-large). Override via EMBEDDING_PROMPT_STYLE when
+        # using a different model family (e.g. nomic, bge).
+        self.prompt_style = _resolve_prompt_style(prompt_style, "e5")
 
         logger.info(
             f"Local Embedder initialized: model={self.model} "
-            f"base_url={self.base_url} dimension={self.dimension}"
+            f"base_url={self.base_url} dimension={self.dimension} "
+            f"prompt_style={self.prompt_style}"
         )