lmnr 0.7.14__tar.gz → 0.7.16__tar.gz

{lmnr-0.7.14 → lmnr-0.7.16}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: lmnr
-Version: 0.7.14
+Version: 0.7.16
 Summary: Python SDK for Laminar
 Author: lmnr.ai
 Author-email: lmnr.ai <founders@lmnr.ai>

{lmnr-0.7.14 → lmnr-0.7.16}/pyproject.toml

@@ -6,7 +6,7 @@
 
 [project]
 name = "lmnr"
-version = "0.7.14"
+version = "0.7.16"
 description = "Python SDK for Laminar"
 authors = [
   { name = "lmnr.ai", email = "founders@lmnr.ai" }

{lmnr-0.7.14 → lmnr-0.7.16}/src/lmnr/opentelemetry_lib/decorators/__init__.py

@@ -225,7 +225,7 @@ def observe_base(
                 detach_context(isolated_ctx_token)
             # span will be ended in the generator
             if isinstance(res, types.GeneratorType):
-                return _handle_generator(span, ctx_token, res)
+                return _handle_generator(span, wrapper, res)
             if isinstance(res, types.AsyncGeneratorType):
                 # async def foo() -> AsyncGenerator[int, None]:
                 # is not considered async in a classical sense in Python,
@@ -234,7 +234,7 @@ def observe_base(
                 # Flags are listed from LSB here:
                 # https://docs.python.org/3/library/inspect.html#inspect-module-co-flags
                 # See also: https://groups.google.com/g/python-tulip/c/6rWweGXLutU?pli=1
-                return _ahandle_generator(span, ctx_token, res)
+                return _ahandle_generator(span, wrapper, res)
 
             _process_output(span, res, ignore_output, output_formatter)
             _cleanup_span(span, wrapper)
@@ -306,7 +306,7 @@ def async_observe_base(
             if isinstance(res, types.AsyncGeneratorType):
                 # probably unreachable, read the comment in the similar
                 # part of the sync wrapper.
-                return await _ahandle_generator(span, ctx_token, res)
+                return await _ahandle_generator(span, wrapper, res)
 
             _process_output(span, res, ignore_output, output_formatter)
             _cleanup_span(span, wrapper)
@@ -317,18 +317,37 @@ def async_observe_base(
     return decorate
 
 
-def _handle_generator(span: Span, wrapper: TracerWrapper, res: Generator):
+def _handle_generator(
+    span: Span,
+    wrapper: TracerWrapper,
+    res: Generator,
+    ignore_output: bool = False,
+    output_formatter: Callable[..., str] | None = None,
+):
+    results = []
     try:
-        yield from res
+        for part in res:
+            results.append(part)
+            yield part
     finally:
+        _process_output(span, results, ignore_output, output_formatter)
         _cleanup_span(span, wrapper)
 
 
-async def _ahandle_generator(span: Span, wrapper: TracerWrapper, res: AsyncGenerator):
+async def _ahandle_generator(
+    span: Span,
+    wrapper: TracerWrapper,
+    res: AsyncGenerator,
+    ignore_output: bool = False,
+    output_formatter: Callable[..., str] | None = None,
+):
+    results = []
     try:
         async for part in res:
+            results.append(part)
             yield part
     finally:
+        _process_output(span, results, ignore_output, output_formatter)
         _cleanup_span(span, wrapper)
 
 

{lmnr-0.7.14 → lmnr-0.7.16}/src/lmnr/opentelemetry_lib/litellm/__init__.py

@@ -406,7 +406,15 @@ try:
                         details.get("cached_tokens"),
                     )
                 # TODO: add audio/image/text token details
-            # TODO: add completion tokens details (reasoning tokens)
+            if usage_dict.get("completion_tokens_details"):
+                details = usage_dict.get("completion_tokens_details", {})
+                details = model_as_dict(details)
+                if details.get("reasoning_tokens"):
+                    set_span_attribute(
+                        span,
+                        "gen_ai.usage.reasoning_tokens",
+                        details.get("reasoning_tokens"),
+                    )
 
         def _process_tool_calls(self, span, tool_calls, choice_index, is_response=True):
             """Process and set tool call attributes on the span"""
@@ -467,17 +475,56 @@ try:
                 content = message.get("content", "")
                 if content is None:
                     continue
+                reasoning_content = message.get("reasoning_content")
+                if reasoning_content:
+                    if isinstance(reasoning_content, str):
+                        reasoning_content = [
+                            {
+                                "type": "text",
+                                "text": reasoning_content,
+                            }
+                        ]
+                    elif not isinstance(reasoning_content, list):
+                        reasoning_content = [
+                            {
+                                "type": "text",
+                                "text": str(reasoning_content),
+                            }
+                        ]
+                else:
+                    reasoning_content = []
                 if isinstance(content, str):
-                    set_span_attribute(span, f"gen_ai.completion.{i}.content", content)
+                    if reasoning_content:
+                        set_span_attribute(
+                            span,
+                            f"gen_ai.completion.{i}.content",
+                            json.dumps(
+                                reasoning_content
+                                + [
+                                    {
+                                        "type": "text",
+                                        "text": content,
+                                    }
+                                ]
+                            ),
+                        )
+                    else:
+                        set_span_attribute(
+                            span,
+                            f"gen_ai.completion.{i}.content",
+                            content,
+                        )
                 elif isinstance(content, list):
                     set_span_attribute(
-                        span, f"gen_ai.completion.{i}.content", json.dumps(content)
+                        span,
+                        f"gen_ai.completion.{i}.content",
+                        json.dumps(reasoning_content + content),
                     )
                 else:
                     set_span_attribute(
                         span,
                         f"gen_ai.completion.{i}.content",
-                        json.dumps(model_as_dict(content)),
+                        json.dumps(reasoning_content + [model_as_dict(content)]),
                     )
 
         def _process_content_part(self, content_part: dict) -> dict:

{lmnr-0.7.14 → lmnr-0.7.16}/src/lmnr/opentelemetry_lib/opentelemetry/instrumentation/google_genai/__init__.py

@@ -272,6 +272,16 @@ def _set_response_attributes(span, response: types.GenerateContentResponse):
 
     if response.usage_metadata:
         usage_dict = to_dict(response.usage_metadata)
+        candidates_token_count = usage_dict.get("candidates_token_count")
+        # unlike OpenAI, and unlike input cached tokens, thinking tokens are
+        # not counted as part of candidates token count, so we need to add them
+        # separately for consistency with other instrumentations
+        thoughts_token_count = usage_dict.get("thoughts_token_count")
+        output_token_count = (
+            (candidates_token_count or 0) + (thoughts_token_count or 0)
+            if candidates_token_count is not None or thoughts_token_count is not None
+            else None
+        )
         set_span_attribute(
             span,
             gen_ai_attributes.GEN_AI_USAGE_INPUT_TOKENS,
@@ -280,7 +290,7 @@ def _set_response_attributes(span, response: types.GenerateContentResponse):
         set_span_attribute(
             span,
             gen_ai_attributes.GEN_AI_USAGE_OUTPUT_TOKENS,
-            usage_dict.get("candidates_token_count"),
+            output_token_count,
         )
         set_span_attribute(
             span,
@@ -292,6 +302,11 @@ def _set_response_attributes(span, response: types.GenerateContentResponse):
             SpanAttributes.LLM_USAGE_CACHE_READ_INPUT_TOKENS,
             usage_dict.get("cached_content_token_count"),
         )
+        set_span_attribute(
+            span,
+            SpanAttributes.LLM_USAGE_REASONING_TOKENS,
+            thoughts_token_count,
+        )
 
     if should_send_prompts():
         set_span_attribute(

{lmnr-0.7.14 → lmnr-0.7.16}/src/lmnr/version.py

@@ -3,7 +3,7 @@ import httpx
 from packaging import version
 
 
-__version__ = "0.7.14"
+__version__ = "0.7.16"
 PYTHON_VERSION = f"{sys.version_info.major}.{sys.version_info.minor}"