Support NVIDIA and Mistral provider setup

mp1/pluto/dispatcher.py

@@ -28,6 +28,10 @@ NVIDIA_RERANK_URLS = (
     "https://ai.api.nvidia.com/v1/retrieval/nvidia/reranking",
     "https://integrate.api.nvidia.com/v1/ranking",
 )
+NVIDIA_MAX_RETRIES_WITH_FALLBACK = 2
+NVIDIA_MAX_RETRIES_DEFAULT = 6
+NVIDIA_TIMEOUT_WITH_FALLBACK = (10, 45)
+NVIDIA_TIMEOUT_DEFAULT = (20, 120)
 
 # Groq has strict RPM limits; semaphore prevents parallel-call overflow.
 _groq_semaphore = threading.BoundedSemaphore(2)
@@ -234,6 +238,9 @@ def _call_nvidia(cfg: ModeConfig, prompt: str) -> str:
     env_var, api_key = _resolve_nvidia_api_key(cfg.model_id)
     if not api_key:
         raise ValueError(f"{env_var} or NVIDIA_API_KEY not set")
+    has_fallback = bool(os.getenv("GROQ_API_KEY", "") or os.getenv("MISTRAL_API_KEY", ""))
+    max_retries = NVIDIA_MAX_RETRIES_WITH_FALLBACK if has_fallback else NVIDIA_MAX_RETRIES_DEFAULT
+    request_timeout = NVIDIA_TIMEOUT_WITH_FALLBACK if has_fallback else NVIDIA_TIMEOUT_DEFAULT
 
     prefix = str(prompt)[:120]
     use_reasoning = any(
@@ -250,7 +257,7 @@ def _call_nvidia(cfg: ModeConfig, prompt: str) -> str:
     if "nano" in cfg.model_id and use_reasoning:
         payload["nvidia"] = {"reasoning": True}
 
-    for attempt in range(6):
+    for attempt in range(max_retries):
         try:
             response = requests.post(
                 NVIDIA_CHAT_URL,
@@ -259,7 +266,7 @@ def _call_nvidia(cfg: ModeConfig, prompt: str) -> str:
                     "Content-Type": "application/json",
                 },
                 json=payload,
-                timeout=(20, 120),
+                timeout=request_timeout,
             )
             if response.status_code == 200:
                 data = response.json()
@@ -289,7 +296,7 @@ def _call_nvidia(cfg: ModeConfig, prompt: str) -> str:
                     "504",
                 ]
             )
-            if (is_rate or is_transient) and attempt < 5:
+            if (is_rate or is_transient) and attempt < max_retries - 1:
                 delay = min(60.0, 5.0 * (1.5**attempt))
                 reason = "rate-limit" if is_rate else "transient error"
                 print(f"  [RETRY] NVIDIA {reason} — waiting {delay:.1f}s (attempt {attempt + 1})")

mp1/pluto/modes.py

@@ -32,7 +32,7 @@ class ModeConfig:
     temperature: float
     max_tokens: int
     compute_profile: str
-    provider: str  # "groq" | "mistral"
+    provider: str  # "nvidia" | "groq" | "mistral"
 
     def to_log_dict(self) -> dict:
         return {
@@ -58,7 +58,7 @@ def _build_registry() -> dict[str, ModeConfig]:
     Embedding + reranking are handled separately in embedder.py and dispatcher.py
     (they use /v1/embeddings and scoring endpoints, not chat completions).
 
-    Fallback: if NVIDIA_API_KEY absent, fall back to Groq (same model sizes).
+    Fallback: if NVIDIA_API_KEY absent, fall back to Groq or Mistral.
     """
     # Check for any NVIDIA key
     nvidia_keys = [
@@ -68,6 +68,7 @@ def _build_registry() -> dict[str, ModeConfig]:
     ]
     nvidia_ready = any(os.getenv(k) for k in nvidia_keys)
     groq_key = os.getenv("GROQ_API_KEY", "").strip()
+    mistral_key = os.getenv("MISTRAL_API_KEY", "").strip()
 
     if nvidia_ready:
         return {
@@ -157,9 +158,57 @@ def _build_registry() -> dict[str, ModeConfig]:
                 provider="groq",
             ),
         }
+    if mistral_key:
+        return _build_mistral_registry()
     return _build_unconfigured_registry()
 
 
+def _build_mistral_registry() -> dict[str, ModeConfig]:
+    """Use Mistral for every mode when it is the only configured chat provider."""
+    return {
+        "MODE_QUICK": ModeConfig(
+            mode_name="MODE_QUICK",
+            model_id="mistral-small-latest",
+            temperature=0.1,
+            max_tokens=1024,
+            compute_profile="fallback",
+            provider="mistral",
+        ),
+        "MODE_REASONING": ModeConfig(
+            mode_name="MODE_REASONING",
+            model_id="mistral-small-latest",
+            temperature=0.3,
+            max_tokens=4096,
+            compute_profile="fallback",
+            provider="mistral",
+        ),
+        "MODE_VISION": ModeConfig(
+            mode_name="MODE_VISION",
+            model_id="mistral-small-latest",
+            temperature=0.1,
+            max_tokens=4096,
+            compute_profile="fallback",
+            provider="mistral",
+        ),
+        "MODE_ULTRA": ModeConfig(
+            mode_name="MODE_ULTRA",
+            model_id="mistral-small-latest",
+            temperature=0.2,
+            max_tokens=4096,
+            compute_profile="fallback",
+            provider="mistral",
+        ),
+        "MODE_GEMINI": ModeConfig(
+            mode_name="MODE_GEMINI",
+            model_id="mistral-small-latest",
+            temperature=0.0,
+            max_tokens=4096,
+            compute_profile="fallback",
+            provider="mistral",
+        ),
+    }
+
+
 def _build_unconfigured_registry() -> dict[str, ModeConfig]:
     """Return placeholder modes so imports work without provider credentials."""
     return {
@@ -210,7 +259,7 @@ MODE_REGISTRY: dict[str, ModeConfig] = _build_registry()
 
 
 def _missing_provider_error() -> EnvironmentError:
-    return EnvironmentError("Neither NVIDIA_API_KEY nor GROQ_API_KEY is set.")
+    return EnvironmentError("None of NVIDIA_API_KEY, GROQ_API_KEY, or MISTRAL_API_KEY is set.")
 
 
 def _is_unconfigured() -> bool: