feat: backend embedding configurabile — EMBED_API=ollama|llamacpp

- llama.cpp: /v1/embeddings OpenAI-compatible con Bearer key (EMBED_API_KEY)
- ollama (default): /api/embed invariato
- EMBED_URL sostituisce OLLAMA_URL (alias retrocompatibile)

gateway/main.py

@@ -43,8 +43,11 @@ from qdrant_client.http import models as qm
 # ---------------------------------------------------------------------------
 QDRANT_URL = os.environ.get("QDRANT_URL", "http://127.0.0.1:6333")
 QDRANT_API_KEY = os.environ.get("QDRANT_API_KEY", "")
-OLLAMA_URL = os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434")
+# Backend embedding: ollama (default) | llamacpp (OpenAI-compatible /v1/embeddings)
+EMBED_API = os.environ.get("EMBED_API", "ollama")
+EMBED_URL = os.environ.get("EMBED_URL", os.environ.get("OLLAMA_URL", "http://127.0.0.1:11434"))
 EMBED_MODEL = os.environ.get("EMBED_MODEL", "bge-m3")
+EMBED_API_KEY = os.environ.get("EMBED_API_KEY", "")
 EMBED_DIM = int(os.environ.get("EMBED_DIM", "1024"))
 COLLECTION = os.environ.get("COLLECTION", "memories")
 # Chiavi condivise (separate da virgola): accesso completo in lettura/scrittura
@@ -392,8 +395,21 @@ def _get_http() -> httpx.AsyncClient:
 
 
 async def embed(text: str) -> list[float]:
+    if EMBED_API == "llamacpp":
+        # llama.cpp: OpenAI-compatible /v1/embeddings
+        headers = {"Content-Type": "application/json"}
+        if EMBED_API_KEY:
+            headers["Authorization"] = f"Bearer {EMBED_API_KEY}"
+        r = await _get_http().post(
+            f"{EMBED_URL}/v1/embeddings",
+            json={"model": EMBED_MODEL, "input": text},
+            headers=headers,
+        )
+        r.raise_for_status()
+        return r.json()["data"][0]["embedding"]
+    # Ollama (default)
     r = await _get_http().post(
-        f"{OLLAMA_URL}/api/embed",
+        f"{EMBED_URL}/api/embed",
         json={"model": EMBED_MODEL, "input": text},
     )
     r.raise_for_status()