fix: DeepSeek pricing + correct vendor attribution in OpenAI instrumentor

pricing.py:
- Add deepseek vendor to MODEL_PRICING (deepseek-chat, deepseek-reasoner, deepseek-coder)
- Add deepseek to DEFAULT_PRICING fallback
- Add DeepSeek fuzzy matching to normalize_model_name()

openai_instr.py:
- Add _detect_vendor() to identify DeepSeek calls by model prefix
- OpenAICostAdapter.calculate_cost() now uses correct vendor pricing per call
- Span name, llm.vendor and llm.system now reflect actual provider (openai or deepseek)
- No separate instrumentor needed — DeepSeek uses OpenAI SDK, patch fires automatically
- OpenAI calls unchanged in behavior, DeepSeek calls no longer misattributed

Author: Meridian

kalibr/instrumentation/openai_instr.py

@@ -17,33 +17,54 @@
 from .base import BaseCostAdapter, BaseInstrumentation
 
 
+def _detect_vendor(model: str) -> str:
+    """Detect the actual vendor from model name.
+
+    The OpenAI Python SDK is used by multiple providers (OpenAI, DeepSeek, etc.)
+    via base_url overrides. We detect the real vendor from the model name so spans
+    and cost calculations are attributed correctly.
+
+    Args:
+        model: Model identifier from the API response
+
+    Returns:
+        Vendor string: "openai" or "deepseek"
+    """
+    model_lower = model.lower()
+    if model_lower.startswith("deepseek-"):
+        return "deepseek"
+    return "openai"
+
+
 class OpenAICostAdapter(BaseCostAdapter):
-    """Cost calculation adapter for OpenAI models.
-    
-    Uses centralized pricing from kalibr.pricing module.
+    """Cost calculation adapter for OpenAI-compatible models.
+
+    Handles OpenAI and any provider that uses the OpenAI SDK (e.g. DeepSeek).
+    Vendor is detected from the model name so cost is always attributed correctly.
     """
 
     def get_vendor_name(self) -> str:
-        """Return vendor name for OpenAI."""
+        """Return vendor name — overridden per-call via calculate_cost."""
         return "openai"
 
     def calculate_cost(self, model: str, usage: Dict[str, int]) -> float:
-        """Calculate cost in USD for an OpenAI API call.
-        
+        """Calculate cost using the correct vendor pricing for this model.
+
         Args:
-            model: Model identifier (e.g., "gpt-4o", "gpt-4o-2024-05-13")
+            model: Model identifier (e.g., "gpt-4o", "deepseek-chat")
             usage: Token usage dict with prompt_tokens and completion_tokens
-            
+
         Returns:
             Cost in USD (rounded to 6 decimal places)
         """
-        # Get pricing from centralized module (handles normalization)
-        pricing = self.get_pricing_for_model(model)
+        from kalibr.pricing import get_pricing
+
+        vendor = _detect_vendor(model)
+        pricing, _ = get_pricing(vendor, model)
 
         prompt_tokens = usage.get("prompt_tokens", 0)
         completion_tokens = usage.get("completion_tokens", 0)
 
-        # Calculate cost (pricing is per 1M tokens)
         input_cost = (prompt_tokens / 1_000_000) * pricing["input"]
         output_cost = (completion_tokens / 1_000_000) * pricing["output"]
 
@@ -124,14 +145,17 @@ def wrapper(self_instance, *args, **kwargs):
             # Extract model from kwargs
             model = kwargs.get("model", "unknown")
 
+            # Detect actual vendor — OpenAI SDK is used by multiple providers
+            vendor = _detect_vendor(model)
+
             # Create span with initial attributes
             with self.tracer.start_as_current_span(
-                "openai.chat.completions.create",
+                f"{vendor}.chat.completions.create",
                 kind=SpanKind.CLIENT,
                 attributes={
-                    "llm.vendor": "openai",
+                    "llm.vendor": vendor,
                     "llm.request.model": model,
-                    "llm.system": "openai",
+                    "llm.system": vendor,
                 },
             ) as span:
                 start_time = time.time()
@@ -167,14 +191,17 @@ async def wrapper(self_instance, *args, **kwargs):
             # Extract model from kwargs
             model = kwargs.get("model", "unknown")
 
+            # Detect actual vendor — OpenAI SDK is used by multiple providers
+            vendor = _detect_vendor(model)
+
             # Create span with initial attributes
             with self.tracer.start_as_current_span(
-                "openai.chat.completions.create",
+                f"{vendor}.chat.completions.create",
                 kind=SpanKind.CLIENT,
                 attributes={
-                    "llm.vendor": "openai",
+                    "llm.vendor": vendor,
                     "llm.request.model": model,
-                    "llm.system": "openai",
+                    "llm.system": vendor,
                 },
             ) as span:
                 start_time = time.time()

kalibr/pricing.py

@@ -60,6 +60,12 @@
         "mistralai/mixtral-8x7b-instruct-v0.1": {"input": 0.60, "output": 0.60},
         # Default fallback for unknown HuggingFace models
     },
+    "deepseek": {
+        # Prices in USD per 1M tokens — https://platform.deepseek.com/docs/pricing
+        "deepseek-chat": {"input": 0.27, "output": 1.10},       # DeepSeek-V3
+        "deepseek-reasoner": {"input": 0.55, "output": 2.19},   # DeepSeek-R1
+        "deepseek-coder": {"input": 0.27, "output": 1.10},      # DeepSeek-Coder-V2
+    },
     "google": {
         # Gemini 2.5 models
         "gemini-2.5-pro": {"input": 1.25, "output": 5.00},
@@ -83,6 +89,7 @@
     "anthropic": {"input": 15.00, "output": 75.00},  # Claude 3 Opus pricing
     "google": {"input": 1.25, "output": 5.00},  # Gemini 1.5 Pro pricing
     "huggingface": {"input": 1.00, "output": 1.00},  # Conservative default
+    "deepseek": {"input": 0.55, "output": 2.19},    # DeepSeek-R1 pricing
 }
 
 
@@ -187,6 +194,15 @@ def normalize_model_name(vendor: str, model_name: str) -> str:
         elif "gemini-1.0-pro" in model_lower or "gemini-pro" in model_lower:
             return "gemini-pro"
 
+    # DeepSeek fuzzy matching
+    elif vendor == "deepseek":
+        if "deepseek-reasoner" in model_lower or "deepseek-r1" in model_lower:
+            return "deepseek-reasoner"
+        elif "deepseek-coder" in model_lower:
+            return "deepseek-coder"
+        elif "deepseek-chat" in model_lower or "deepseek-v" in model_lower:
+            return "deepseek-chat"
+
     # Return original if no match found
     return model_lower