feat(stt): mémoire long-terme sémantique via Qdrant (5b) (#10)

Serveur : longterm.py — collection Qdrant stt-memory (embeddings Qwen3 gpu-01, dim auto,
Cosine), recall top-k injecté au prompt, remember des tours user. Tout dégrade proprement
si Qdrant/embeddings injoignables (la mémoire court-terme tient). Env STT_MEMORY_LONGTERM,
STT_QDRANT_URL, STT_EMBED_URL, STT_MEMORY_TOPK.

Testé en process : dégradation OK (Qdrant down → mem=0, pas de crash, court-terme tient).
Qdrant réparé le 17/06 (5c). Recherche sémantique réelle à valider sur cluster.


Claude-Session: https://claude.ai/code/session_013FmcxGsyXZXogiAHQLjnZT

Co-authored-by: Claude <noreply@anthropic.com>

stt/server/stt_server/app.py

@@ -14,10 +14,12 @@ from pydantic import BaseModel
 from stt_server import __version__
 from stt_server.brain import ask as brain_ask
 from stt_server.config import settings
+from stt_server.longterm import LongTermMemory
 from stt_server.memory import SessionStore
 
 app = FastAPI(title="STT-server", version=__version__)
 sessions = SessionStore()
+longterm = LongTermMemory() if settings.memory_longterm else None
 
 
 class AskRequest(BaseModel):
@@ -60,13 +62,16 @@ async def v1_ask(req: AskRequest) -> AskReply:
             detail=f"modèle '{model}' non autorisé ; dispo : {settings.allowed_models}",
         )
     history = sessions.history(req.session_id) if req.session_id else None
+    memories = await longterm.recall(text) if longterm else None
     try:
-        reply = await brain_ask(text, model, history)
+        reply = await brain_ask(text, model, history, memories)
     except httpx.HTTPError as e:
         raise HTTPException(status_code=502, detail=f"upstream LiteLLM : {e}") from e
     if req.session_id:
         sessions.add(req.session_id, "user", text)
         sessions.add(req.session_id, "assistant", reply)
+    if longterm:
+        await longterm.remember(req.session_id or "anon", text)
     return AskReply(reply=reply, model=model)
 
 

stt/server/stt_server/brain.py

@@ -12,8 +12,20 @@ import httpx
 from stt_server.config import settings
 
 
-async def ask(text: str, model: str | None = None, history: list[dict] | None = None) -> str:
-    messages = [{"role": "system", "content": settings.system_prompt}]
+async def ask(
+    text: str,
+    model: str | None = None,
+    history: list[dict] | None = None,
+    memories: list[str] | None = None,
+) -> str:
+    system = settings.system_prompt
+    if memories:
+        souvenirs = "\n".join(f"- {m}" for m in memories)
+        system += (
+            "\n\nÉléments de mémoire long-terme (peuvent aider, ignore si hors-sujet) :\n"
+            + souvenirs
+        )
+    messages = [{"role": "system", "content": system}]
     if history:
         messages += history
     messages.append({"role": "user", "content": text})

stt/server/stt_server/config.py

@@ -32,5 +32,14 @@ class Settings:
     temperature: float = float(os.getenv("STT_TEMPERATURE", "0.7"))
     request_timeout: float = float(os.getenv("STT_REQUEST_TIMEOUT", "60"))
 
+    # Mémoire long-terme (Qdrant) — dégrade proprement si Qdrant/embeddings injoignables
+    memory_longterm: bool = os.getenv("STT_MEMORY_LONGTERM", "true").lower() == "true"
+    qdrant_url: str = os.getenv("STT_QDRANT_URL", "http://192.168.10.1:6333")
+    qdrant_collection: str = os.getenv("STT_QDRANT_COLLECTION", "stt-memory")
+    # Embeddings : Qwen3 sur llama-server gpu-01 (comme le RAG). dim 4096.
+    embed_url: str = os.getenv("STT_EMBED_URL", "http://192.168.10.20:1234/v1/embeddings")
+    embed_model: str = os.getenv("STT_EMBED_MODEL", "qwen3-8b")
+    memory_top_k: int = int(os.getenv("STT_MEMORY_TOPK", "3"))
+
 
 settings = Settings()

stt/server/stt_server/longterm.py

@@ -0,0 +1,91 @@
+"""Mémoire long-terme sémantique (Qdrant).
+
+Stocke les tours utilisateur comme vecteurs dans la collection `stt-memory` et retrouve
+les souvenirs pertinents pour les injecter dans le prompt. Embeddings via Qwen3 (llama-server
+gpu-01), comme le RAG. **Tout dégrade proprement** : si Qdrant ou l'endpoint d'embedding
+est injoignable, `recall` renvoie [] et `remember` ne fait rien (la mémoire court-terme
+de session continue de fonctionner).
+
+> Caveat qualité : Qwen3 n'est pas un modèle d'embedding dédié (cosinus uniformément hauts) ;
+> la recherche est approximative. Voir admin/ia/rag.md pour la piste nomic-embed-text.
+"""
+
+from __future__ import annotations
+
+import time
+import uuid
+
+import httpx
+
+from stt_server.config import settings
+
+
+class LongTermMemory:
+    def __init__(self) -> None:
+        self.qdrant = settings.qdrant_url.rstrip("/")
+        self.collection = settings.qdrant_collection
+        self.embed_url = settings.embed_url
+        self.embed_model = settings.embed_model
+        self.top_k = settings.memory_top_k
+        self._ready = False
+
+    async def _embed(self, client: httpx.AsyncClient, text: str) -> list[float]:
+        r = await client.post(
+            self.embed_url,
+            json={"model": self.embed_model, "input": text},
+            timeout=30,
+        )
+        r.raise_for_status()
+        return r.json()["data"][0]["embedding"]
+
+    async def _ensure_collection(self, client: httpx.AsyncClient, dim: int) -> None:
+        if self._ready:
+            return
+        r = await client.get(f"{self.qdrant}/collections/{self.collection}")
+        if r.status_code == 404:
+            await client.put(
+                f"{self.qdrant}/collections/{self.collection}",
+                json={"vectors": {"size": dim, "distance": "Cosine"}},
+            )
+        self._ready = True
+
+    async def recall(self, text: str) -> list[str]:
+        """Souvenirs pertinents (texte) ou [] si indisponible."""
+        try:
+            async with httpx.AsyncClient(timeout=20) as client:
+                vec = await self._embed(client, text)
+                r = await client.post(
+                    f"{self.qdrant}/collections/{self.collection}/points/search",
+                    json={"vector": vec, "limit": self.top_k, "with_payload": True},
+                )
+                if r.status_code == 404:  # collection pas encore créée
+                    return []
+                r.raise_for_status()
+                pts = r.json().get("result", [])
+                return [p["payload"]["text"] for p in pts if p.get("payload", {}).get("text")]
+        except Exception:  # noqa: BLE001 — dégrade silencieusement
+            return []
+
+    async def remember(self, session_id: str, text: str) -> None:
+        try:
+            async with httpx.AsyncClient(timeout=20) as client:
+                vec = await self._embed(client, text)
+                await self._ensure_collection(client, len(vec))
+                await client.put(
+                    f"{self.qdrant}/collections/{self.collection}/points",
+                    json={
+                        "points": [
+                            {
+                                "id": str(uuid.uuid4()),
+                                "vector": vec,
+                                "payload": {
+                                    "text": text,
+                                    "session_id": session_id,
+                                    "ts": time.time(),
+                                },
+                            }
+                        ]
+                    },
+                )
+        except Exception:  # noqa: BLE001 — dégrade silencieusement
+            pass