diff --git a/apps/api/src/services/nvidia_provider.py b/apps/api/src/services/nvidia_provider.py
index 7fa80f57..014395dc 100644
--- a/apps/api/src/services/nvidia_provider.py
+++ b/apps/api/src/services/nvidia_provider.py
@@ -491,23 +491,27 @@ class NvidiaProvider:
                 span.set_attribute("ai.total_tokens", total_tokens)
 
                 # P1-1: Langfuse Generation 記錄
-                langfuse_ctx.generation(
-                    name="nvidia_nemotron",
-                    model=model,
-                    input={"messages": messages, "tools": tool_names},
-                    output={
-                        "tool_calls": [
-                            {"name": tc.tool_name, "args": tc.arguments}
-                            for tc in tool_calls
-                            if tc.valid
-                        ]
-                    },
-                    usage={"input": prompt_tokens, "output": completion_tokens},
-                    metadata={
-                        "latency_ms": round(latency_ms, 2),
-                        "valid_count": sum(1 for tc in tool_calls if tc.valid),
-                    },
-                )
+                if langfuse_ctx and hasattr(langfuse_ctx, "generation"):
+                    try:
+                        langfuse_ctx.generation(
+                            name="nvidia_nemotron",
+                            model=model,
+                            input={"messages": messages, "tools": tool_names},
+                            output={
+                                "tool_calls": [
+                                    {"name": tc.tool_name, "args": tc.arguments}
+                                    for tc in tool_calls
+                                    if tc.valid
+                                ]
+                            },
+                            usage={"input": prompt_tokens, "output": completion_tokens},
+                            metadata={
+                                "latency_ms": round(latency_ms, 2),
+                                "valid_count": sum(1 for tc in tool_calls if tc.valid),
+                            },
+                        )
+                    except Exception as le:
+                        logger.warning("langfuse_generation_failed_safe", error=str(le))
 
                 # P3-3: Prometheus 成功指標
                 for tc in tool_calls:
@@ -750,17 +754,21 @@ class NvidiaProvider:
                     NVIDIA_LATENCY_HISTOGRAM.observe(latency_ms / 1000)
 
                     # Langfuse
-                    langfuse_ctx.generation(
-                        name="nvidia_chat",
-                        model=model_name,
-                        input=prompt[:500],
-                        output=text[:500],
-                        metadata={
-                            "total_tokens": total_tokens,
-                            "cost_usd": cost_usd,
-                            "latency_ms": round(latency_ms, 2),
-                        },
-                    )
+                    if langfuse_ctx and hasattr(langfuse_ctx, "generation"):
+                        try:
+                            langfuse_ctx.generation(
+                                name="nvidia_chat",
+                                model=model_name,
+                                input=prompt[:500],
+                                output=text[:500],
+                                metadata={
+                                    "total_tokens": total_tokens,
+                                    "cost_usd": cost_usd,
+                                    "latency_ms": round(latency_ms, 2),
+                                },
+                            )
+                        except Exception as le:
+                            logger.warning("langfuse_chat_generation_failed_safe", error=str(le))
 
                     logger.info(
                         "nvidia_chat_response_received",
@@ -797,7 +805,13 @@ class NvidiaProvider:
                 except Exception as e:
                     self._circuit_breaker.record_failure()
                     NVIDIA_REQUESTS_TOTAL.labels(status="error", tool_name="chat").inc()
-                    logger.warning("nvidia_chat_failed", error=str(e), error_type=type(e).__name__)
+                    import traceback
+                    logger.warning(
+                        "nvidia_chat_failed", 
+                        error=str(e), 
+                        error_type=type(e).__name__,
+                        stacktrace=traceback.format_exc()
+                    )
                     return str(e), False, 0, 0.0