🔨 Migrate Langfuse to v3.x and update Gemini model names

.cursor/rules/langfuse.mdc

@@ -3,17 +3,28 @@ description:
 globs: *.py
 alwaysApply: false
 ---
-LangFuse is an LLM observability tool, to see how LLMs are behaving in our application. It uses callbacks to reord LLM inference. You should always use it when using LLM applications, to track if the LLM is behaving as we want it to.
+Langfuse is an LLM observability tool used to monitor LLM behavior in our application. It uses decorators and callbacks to record LLM inference. Always use it in LLM-related modules to ensure traceability.
 
-Below is the most typical usage.
+### Langfuse v3 Migration Guide
+As of Langfuse SDK v3.x, the `langfuse.decorators` module structure has changed.
+
+**Typical Usage:**
 
 ```python
-from langfuse.decorators import observe, langfuse_context
+from langfuse import observe, get_client
 
-@observe
+@observe()
 def function_name(...):
+    # To rename the span/observation or update metadata:
+    get_client().update_current_span(name=f"descriptive_name_{id}")
 
-    # If we don't want the name of the span to be "function_name", and want to give a more descriptive name like with email, we should use the following
-    langfuse_context.update_current_observation(name=f"{email}")
+    # To get trace or observation IDs:
+    trace_id = get_client().get_current_trace_id()
+    observation_id = get_client().get_current_observation_id()
 ```
 
+**Key Changes from v2:**
+1. Import `observe` and `get_client` directly from `langfuse`.
+2. Replace `langfuse_context` with `get_client()`.
+3. Use `update_current_span()` instead of `update_current_observation()`.
+

common/global_config.yaml

@@ -1,5 +1,6 @@
-model_name: gemini/gemini-3-flash
+model_name: gemini/gemini-3-flash-preview
 dot_global_config_health_check: true
+DEV_ENV: dev
 
 example_parent:
   example_child: "example_value"
@@ -8,8 +9,8 @@ example_parent:
 # LLMs
 ########################################################
 default_llm:
-  default_model: gemini/gemini-3-flash
-  fallback_model: gemini/gemini-2.5-flash
+  default_model: gemini/gemini-3-flash-preview
+  fallback_model: gemini/gemini-2.5-flash-preview
   default_temperature: 0.5
   default_max_tokens: 100000
 

utils/llm/dspy_inference.py

@@ -2,7 +2,7 @@
 from typing import Any
 
 import dspy
-from langfuse.decorators import observe  # type: ignore
+from langfuse import observe
 from litellm.exceptions import RateLimitError, ServiceUnavailableError
 from loguru import logger as log
 from tenacity import (

utils/llm/dspy_langfuse.py

@@ -4,8 +4,7 @@
 from dspy.adapters import Image as dspy_Image
 from dspy.signatures import Signature as dspy_Signature
 from dspy.utils.callback import BaseCallback
-from langfuse.client import Langfuse, StatefulGenerationClient  # type: ignore
-from langfuse.decorators import langfuse_context  # type: ignore
+from langfuse import Langfuse, LangfuseGeneration, get_client
 from litellm.cost_calculator import completion_cost
 from loguru import logger as log
 from pydantic import BaseModel, Field, ValidationError
@@ -52,7 +51,7 @@ def __init__(self, signature: type[dspy_Signature]) -> None:
         )
         self.current_prompt = contextvars.ContextVar[str]("current_prompt")
         self.current_completion = contextvars.ContextVar[str]("current_completion")
-        self.current_span = contextvars.ContextVar[StatefulGenerationClient | None](
+        self.current_span = contextvars.ContextVar[LangfuseGeneration | None](
             "current_span"
         )
         self.model_name_at_span_creation = contextvars.ContextVar[str | None](
@@ -91,8 +90,8 @@ def on_module_end(  # noqa
         exception: Exception | None = None,  # noqa
     ) -> None:
         metadata = {
-            "existing_trace_id": langfuse_context.get_current_trace_id(),
-            "parent_observation_id": langfuse_context.get_current_observation_id(),
+            "existing_trace_id": get_client().get_current_trace_id(),
+            "parent_observation_id": get_client().get_current_observation_id(),
         }
         outputs_extracted = {}  # Default to empty dict
         if outputs is not None:
@@ -102,7 +101,7 @@ def on_module_end(  # noqa
                 outputs_extracted = {"value": outputs}
             except Exception as e:
                 outputs_extracted = {"error_extracting_module_output": str(e)}
-        langfuse_context.update_current_observation(
+        get_client().update_current_span(
             input=self.input_field_values.get(None) or {},
             output=outputs_extracted,
             metadata=metadata,
@@ -134,9 +133,9 @@ def on_lm_start(  # noqa
         self.current_system_prompt.set(system_prompt)
         self.current_prompt.set(user_input)
         self.model_name_at_span_creation.set(model_name)
-        trace_id = langfuse_context.get_current_trace_id()
-        parent_observation_id = langfuse_context.get_current_observation_id()
-        span_obj: StatefulGenerationClient | None = None
+        trace_id = get_client().get_current_trace_id()
+        parent_observation_id = get_client().get_current_observation_id()
+        span_obj: LangfuseGeneration | None = None
         if trace_id:
             span_obj = self.langfuse.generation(
                 input=user_input,
@@ -392,8 +391,8 @@ def on_tool_start(  # noqa
 
         log.debug(f"Tool call started: {tool_name} with args: {tool_args}")
 
-        trace_id = langfuse_context.get_current_trace_id()
-        parent_observation_id = langfuse_context.get_current_observation_id()
+        trace_id = get_client().get_current_trace_id()
+        parent_observation_id = get_client().get_current_observation_id()
 
         if trace_id:
             # Create a span for the tool call