feat: switch deck generation from Perplexity to local Ollama (gpt-oss:20b)

coach/clients/perplexity_client.py

@@ -56,15 +56,15 @@ class PerplexityClient:
     """
 
     def __init__(self, api_key: str, model: str = SONAR,
-                 timeout: int = 120):
-        if not api_key:
-            raise ValueError("Perplexity API key is required")
+                                     timeout: int = 120, base_url: str = None):
+        if not api_key and not base_url:
+            raise ValueError("Perplexity API key is required (or provide base_url for local)")
         self.api_key = api_key
         self.model = model
         self.timeout = timeout
         self._client = OpenAI(
             api_key=api_key,
-            base_url="https://api.perplexity.ai",
+            base_url=base_url or "https://api.perplexity.ai",
             timeout=timeout,
         )
 

coach/config.py

@@ -13,8 +13,8 @@ from pathlib import Path
 LAB_ROOT = Path(os.environ.get("COMMANDER_LAB_ROOT", Path(__file__).parent.parent))
 
 # ── LM Studio (OpenAI-compatible API) ──────────────────────
-LM_STUDIO_URL = os.environ.get("LM_STUDIO_URL", "http://192.168.0.241:1234/v1")
-LM_STUDIO_MODEL = os.environ.get("LM_STUDIO_MODEL", "local-model")
+LM_STUDIO_URL = os.environ.get("LM_STUDIO_URL", "http://localhost:11434/v1")
+LM_STUDIO_MODEL = os.environ.get("LM_STUDIO_MODEL", "gpt-oss:20b")
 LM_STUDIO_TIMEOUT = int(os.environ.get("LM_STUDIO_TIMEOUT", "120"))  # seconds
 LM_STUDIO_MAX_RETRIES = 3
 
@@ -41,9 +41,10 @@ DEFAULT_MAX_TOKENS = 8192
 
 # ── Perplexity Deck Generation Settings ───────────────────────
 # Provider for deck generation: "perplexity" (V3) or "local" (V2 collection-based)
-DECK_GEN_PROVIDER = os.environ.get("DECK_GEN_PROVIDER", "perplexity")
+DECK_GEN_PROVIDER = os.environ.get("DECK_GEN_PROVIDER", "local")
 # Model for deck generation (sonar = fast/$0.004, sonar-pro = deep/$0.04)
-DECK_GEN_MODEL = os.environ.get("DECK_GEN_MODEL", "sonar")
+DECK_GEN_MODEL = os.environ.get("DECK_GEN_MODEL", "gpt-oss:20b")
+DECK_GEN_BASE_URL = os.environ.get("DECK_GEN_BASE_URL", "http://localhost:11434/v1")
 DECK_GEN_TEMPERATURE = float(os.environ.get("DECK_GEN_TEMPERATURE", "0.2"))
 DECK_GEN_MAX_TOKENS = int(os.environ.get("DECK_GEN_MAX_TOKENS", "16384"))
 
@@ -53,7 +54,7 @@ SUBSTITUTION_MIN_SIMILARITY = float(os.environ.get("SUBSTITUTION_MIN_SIMILARITY"
 # Maximum number of alternatives to suggest per missing card
 SUBSTITUTION_MAX_ALTERNATIVES = int(os.environ.get("SUBSTITUTION_MAX_ALTERNATIVES", "5"))
 # Model for substitution fallback (sonar is fine — small focused queries)
-SUBSTITUTION_MODEL = os.environ.get("SUBSTITUTION_MODEL", "sonar")
+SUBSTITUTION_MODEL = os.environ.get("SUBSTITUTION_MODEL", "gpt-oss:20b")
 # Enable Perplexity fallback for low-confidence embedding matches
 SUBSTITUTION_USE_PPLX_FALLBACK = os.environ.get("SUBSTITUTION_USE_PPLX_FALLBACK", "true").lower() == "true"
 

routes/coach.py

@@ -91,18 +91,24 @@ def init_coach_service():
 
         log_coach.info("  Coach:        Service initialized")
 
-        # Initialize V3 Deck Generator (Perplexity)
+        if CFG.pplx_api_key or os.environ.get('DECK_GEN_PROVIDER') == 'local':
         _deck_gen_v3_error = None
-        if CFG.pplx_api_key:
             try:
                 from coach.clients.perplexity_client import PerplexityClient
                 from coach.services.deck_generator import DeckGeneratorV3
-                from coach.config import DECK_GEN_MODEL
+                from coach.config import DECK_GEN_MODEL, DECK_GEN_PROVIDER, DECK_GEN_BASE_URL
 
-                pplx_client = PerplexityClient(
-                    api_key=CFG.pplx_api_key,
-                    model=DECK_GEN_MODEL,
-                )
+                if DECK_GEN_PROVIDER == 'local':
+                    pplx_client = PerplexityClient(
+                        api_key='ollama',
+                        model=DECK_GEN_MODEL,
+                        base_url=DECK_GEN_BASE_URL,
+                    )
+                else:
+                    pplx_client = PerplexityClient(
+                        api_key=CFG.pplx_api_key,
+                        model=DECK_GEN_MODEL,
+                    )
                 _deck_gen_v3 = DeckGeneratorV3(
                     pplx_client=pplx_client,
                     db_conn_factory=_get_db_conn,

routes/deckgen.py

@@ -771,26 +771,33 @@ async def _call_pplx_api(messages: list[dict], max_tokens: int = 4096, temperatu
     Uses httpx.AsyncClient so the uvicorn event loop is never stalled.
     Returns the assistant message content.
     """
-    if not CFG.pplx_api_key:
-        raise HTTPException(400, 'Perplexity API key not configured. Set PPLX_API_KEY env var or --pplx-key.')
+    from coach.config import DECK_GEN_PROVIDER, DECK_GEN_BASE_URL, DECK_GEN_MODEL
+
+    if DECK_GEN_PROVIDER == 'local':
+        base_url = DECK_GEN_BASE_URL
+        model = DECK_GEN_MODEL
+        headers = {'Content-Type': 'application/json'}
+    else:
+        if not CFG.pplx_api_key:
+            raise HTTPException(400, 'Perplexity API key not configured. Set PPLX_API_KEY env var or --pplx-key.')
+        base_url = 'https://api.perplexity.ai'
+        model = 'sonar'
+        headers = {
+            'Content-Type': 'application/json',
+            'Authorization': f'Bearer {CFG.pplx_api_key}',
+        }
     payload = {
-        'model': 'sonar',
+        'model': model,
         'messages': messages,
         'max_tokens': max_tokens,
         'temperature': temperature,
-        'return_related_questions': False,
-    }
-    headers = {
-        'Content-Type': 'application/json',
-        'Authorization': f'Bearer {CFG.pplx_api_key}',
     }
     try:
         async with httpx.AsyncClient(timeout=120.0) as client:
             resp = await client.post(
-                'https://api.perplexity.ai/chat/completions',
+                f'{base_url}/chat/completions',
                 json=payload,
                 headers=headers,
-            )
             resp.raise_for_status()
             data = resp.json()
     except httpx.HTTPStatusError as e: