opentelemetry-instrumentation-openai 0.18.1__tar.gz → 0.19.0__tar.gz

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.1
 Name: opentelemetry-instrumentation-openai
-Version: 0.18.1
+Version: 0.19.0
 Summary: OpenTelemetry OpenAI instrumentation
 Home-page: https://github.com/traceloop/openllmetry/tree/main/packages/opentelemetry-instrumentation-openai
 License: Apache-2.0

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/opentelemetry/instrumentation/openai/shared/__init__.py

@@ -112,7 +112,9 @@ def _set_request_attributes(span, kwargs):
     _set_span_attribute(
         span, SpanAttributes.LLM_REQUEST_MAX_TOKENS, kwargs.get("max_tokens")
     )
-    _set_span_attribute(span, SpanAttributes.LLM_REQUEST_TEMPERATURE, kwargs.get("temperature"))
+    _set_span_attribute(
+        span, SpanAttributes.LLM_REQUEST_TEMPERATURE, kwargs.get("temperature")
+    )
     _set_span_attribute(span, SpanAttributes.LLM_REQUEST_TOP_P, kwargs.get("top_p"))
     _set_span_attribute(
         span, SpanAttributes.LLM_FREQUENCY_PENALTY, kwargs.get("frequency_penalty")
@@ -139,6 +141,10 @@ def _set_response_attributes(span, response):
 
     _set_span_attribute(span, SpanAttributes.LLM_RESPONSE_MODEL, response.get("model"))
 
+    _set_span_attribute(
+        span, "gen_ai.openai.system_fingerprint", response.get("system_fingerprint")
+    )
+
     usage = response.get("usage")
     if not usage:
         return

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/opentelemetry/instrumentation/openai/shared/chat_wrappers.py

@@ -87,18 +87,32 @@ def chat_wrapper(
 
     if is_streaming_response(response):
         # span will be closed after the generator is done
-        return ChatStream(
-            span,
-            response,
-            instance,
-            token_counter,
-            choice_counter,
-            duration_histogram,
-            streaming_time_to_first_token,
-            streaming_time_to_generate,
-            start_time,
-            kwargs,
-        )
+        if is_openai_v1():
+            return ChatStream(
+                span,
+                response,
+                instance,
+                token_counter,
+                choice_counter,
+                duration_histogram,
+                streaming_time_to_first_token,
+                streaming_time_to_generate,
+                start_time,
+                kwargs,
+            )
+        else:
+            return _build_from_streaming_response(
+                span,
+                response,
+                instance,
+                token_counter,
+                choice_counter,
+                duration_histogram,
+                streaming_time_to_first_token,
+                streaming_time_to_generate,
+                start_time,
+                kwargs,
+            )
 
     duration = end_time - start_time
 
@@ -161,18 +175,32 @@ async def achat_wrapper(
 
     if is_streaming_response(response):
         # span will be closed after the generator is done
-        return ChatStream(
-            span,
-            response,
-            instance,
-            token_counter,
-            choice_counter,
-            duration_histogram,
-            streaming_time_to_first_token,
-            streaming_time_to_generate,
-            start_time,
-            kwargs,
-        )
+        if is_openai_v1():
+            return ChatStream(
+                span,
+                response,
+                instance,
+                token_counter,
+                choice_counter,
+                duration_histogram,
+                streaming_time_to_first_token,
+                streaming_time_to_generate,
+                start_time,
+                kwargs,
+            )
+        else:
+            return _abuild_from_streaming_response(
+                span,
+                response,
+                instance,
+                token_counter,
+                choice_counter,
+                duration_histogram,
+                streaming_time_to_first_token,
+                streaming_time_to_generate,
+                start_time,
+                kwargs,
+            )
 
     duration = end_time - start_time
 
@@ -277,7 +305,7 @@ def _set_token_counter_metrics(token_counter, usage, shared_attributes):
                 **shared_attributes,
                 "llm.usage.token_type": name.split("_")[0],
             }
-            token_counter.add(val, attributes=attributes_with_token_type)
+            token_counter.record(val, attributes=attributes_with_token_type)
 
 
 def _set_prompts(span, messages):
@@ -359,7 +387,9 @@ def _set_streaming_token_metrics(
         prompt_content = ""
         # setting the default model_name as gpt-4. As this uses the embedding "cl100k_base" that
         # is used by most of the other model.
-        model_name = request_kwargs.get("model") or complete_response.get("model") or "gpt-4"
+        model_name = (
+            request_kwargs.get("model") or complete_response.get("model") or "gpt-4"
+        )
         for msg in request_kwargs.get("messages"):
             if msg.get("content"):
                 prompt_content += msg.get("content")
@@ -392,17 +422,27 @@ def _set_streaming_token_metrics(
                 **shared_attributes,
                 "llm.usage.token_type": "prompt",
             }
-            token_counter.add(prompt_usage, attributes=attributes_with_token_type)
+            token_counter.record(prompt_usage, attributes=attributes_with_token_type)
 
         if type(completion_usage) is int and completion_usage >= 0:
             attributes_with_token_type = {
                 **shared_attributes,
                 "llm.usage.token_type": "completion",
             }
-            token_counter.add(completion_usage, attributes=attributes_with_token_type)
+            token_counter.record(completion_usage, attributes=attributes_with_token_type)
 
 
 class ChatStream(ObjectProxy):
+    _span = None
+    _instance = None
+    _token_counter = None
+    _choice_counter = None
+    _duration_histogram = None
+    _streaming_time_to_first_token = None
+    _streaming_time_to_generate = None
+    _start_time = None
+    _request_kwargs = None
+
     def __init__(
         self,
         span,
@@ -539,3 +579,166 @@ class ChatStream(ObjectProxy):
 
         self._span.set_status(Status(StatusCode.OK))
         self._span.end()
+
+
+# Backward compatibility with OpenAI v0
+
+
+@dont_throw
+def _build_from_streaming_response(
+    span,
+    response,
+    instance=None,
+    token_counter=None,
+    choice_counter=None,
+    duration_histogram=None,
+    streaming_time_to_first_token=None,
+    streaming_time_to_generate=None,
+    start_time=None,
+    request_kwargs=None,
+):
+    complete_response = {"choices": [], "model": ""}
+
+    first_token = True
+    time_of_first_token = start_time  # will be updated when first token is received
+
+    for item in response:
+        span.add_event(name="llm.content.completion.chunk")
+
+        item_to_yield = item
+
+        if first_token and streaming_time_to_first_token:
+            time_of_first_token = time.time()
+            streaming_time_to_first_token.record(time_of_first_token - start_time)
+            first_token = False
+
+        _accumulate_stream_items(item, complete_response)
+
+        yield item_to_yield
+
+    shared_attributes = {
+        "gen_ai.response.model": complete_response.get("model") or None,
+        "server.address": _get_openai_base_url(instance),
+        "stream": True,
+    }
+
+    if not is_azure_openai(instance):
+        _set_streaming_token_metrics(
+            request_kwargs, complete_response, span, token_counter, shared_attributes
+        )
+
+    # choice metrics
+    if choice_counter and complete_response.get("choices"):
+        _set_choice_counter_metrics(
+            choice_counter, complete_response.get("choices"), shared_attributes
+        )
+
+    # duration metrics
+    if start_time and isinstance(start_time, (float, int)):
+        duration = time.time() - start_time
+    else:
+        duration = None
+    if duration and isinstance(duration, (float, int)) and duration_histogram:
+        duration_histogram.record(duration, attributes=shared_attributes)
+    if streaming_time_to_generate and time_of_first_token:
+        streaming_time_to_generate.record(time.time() - time_of_first_token)
+
+    _set_response_attributes(span, complete_response)
+
+    if should_send_prompts():
+        _set_completions(span, complete_response.get("choices"))
+
+    span.set_status(Status(StatusCode.OK))
+    span.end()
+
+
+@dont_throw
+async def _abuild_from_streaming_response(
+    span,
+    response,
+    instance=None,
+    token_counter=None,
+    choice_counter=None,
+    duration_histogram=None,
+    streaming_time_to_first_token=None,
+    streaming_time_to_generate=None,
+    start_time=None,
+    request_kwargs=None,
+):
+    complete_response = {"choices": [], "model": ""}
+
+    first_token = True
+    time_of_first_token = start_time  # will be updated when first token is received
+
+    async for item in response:
+        span.add_event(name="llm.content.completion.chunk")
+
+        item_to_yield = item
+
+        if first_token and streaming_time_to_first_token:
+            time_of_first_token = time.time()
+            streaming_time_to_first_token.record(time_of_first_token - start_time)
+            first_token = False
+
+        _accumulate_stream_items(item, complete_response)
+
+        yield item_to_yield
+
+    shared_attributes = {
+        "gen_ai.response.model": complete_response.get("model") or None,
+        "server.address": _get_openai_base_url(instance),
+        "stream": True,
+    }
+
+    if not is_azure_openai(instance):
+        _set_streaming_token_metrics(
+            request_kwargs, complete_response, span, token_counter, shared_attributes
+        )
+
+    # choice metrics
+    if choice_counter and complete_response.get("choices"):
+        _set_choice_counter_metrics(
+            choice_counter, complete_response.get("choices"), shared_attributes
+        )
+
+    # duration metrics
+    if start_time and isinstance(start_time, (float, int)):
+        duration = time.time() - start_time
+    else:
+        duration = None
+    if duration and isinstance(duration, (float, int)) and duration_histogram:
+        duration_histogram.record(duration, attributes=shared_attributes)
+    if streaming_time_to_generate and time_of_first_token:
+        streaming_time_to_generate.record(time.time() - time_of_first_token)
+
+    _set_response_attributes(span, complete_response)
+
+    if should_send_prompts():
+        _set_completions(span, complete_response.get("choices"))
+
+    span.set_status(Status(StatusCode.OK))
+    span.end()
+
+
+def _accumulate_stream_items(item, complete_response):
+    if is_openai_v1():
+        item = model_as_dict(item)
+
+    complete_response["model"] = item.get("model")
+
+    for choice in item.get("choices"):
+        index = choice.get("index")
+        if len(complete_response.get("choices")) <= index:
+            complete_response["choices"].append(
+                {"index": index, "message": {"content": "", "role": ""}}
+            )
+        complete_choice = complete_response.get("choices")[index]
+        if choice.get("finish_reason"):
+            complete_choice["finish_reason"] = choice.get("finish_reason")
+
+        delta = choice.get("delta")
+
+        if delta and delta.get("content"):
+            complete_choice["message"]["content"] += delta.get("content")
+        if delta and delta.get("role"):
+            complete_choice["message"]["role"] = delta.get("role")

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/opentelemetry/instrumentation/openai/shared/embeddings_wrappers.py

@@ -202,7 +202,7 @@ def _set_embeddings_metrics(
                     **shared_attributes,
                     "llm.usage.token_type": name.split("_")[0],
                 }
-                token_counter.add(val, attributes=attributes_with_token_type)
+                token_counter.record(val, attributes=attributes_with_token_type)
 
     # vec size metrics
     # should use counter for vector_size?

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/opentelemetry/instrumentation/openai/v0/__init__.py

@@ -35,20 +35,20 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
         meter = get_meter(__name__, __version__, meter_provider)
 
         if is_metrics_enabled():
-            chat_token_counter = meter.create_counter(
-                name="llm.openai.chat_completions.tokens",
+            tokens_histogram = meter.create_histogram(
+                name="gen_ai.client.token.usage",
                 unit="token",
                 description="Number of tokens used in prompt and completions",
             )
 
             chat_choice_counter = meter.create_counter(
-                name="llm.openai.chat_completions.choices",
+                name="gen_ai.client.generation.choices",
                 unit="choice",
                 description="Number of choices returned by chat completions call",
             )
 
             chat_duration_histogram = meter.create_histogram(
-                name="llm.openai.chat_completions.duration",
+                name="gen_ai.client.operation.duration",
                 unit="s",
                 description="Duration of chat completion operation",
             )
@@ -71,7 +71,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             )
         else:
             (
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -80,12 +80,6 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             ) = (None, None, None, None, None, None)
 
         if is_metrics_enabled():
-            embeddings_token_counter = meter.create_counter(
-                name="llm.openai.embeddings.tokens",
-                unit="token",
-                description="Number of tokens used in prompt and completions",
-            )
-
             embeddings_vector_size_counter = meter.create_counter(
                 name="llm.openai.embeddings.vector_size",
                 unit="element",
@@ -105,7 +99,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             )
         else:
             (
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,
@@ -120,7 +114,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             "ChatCompletion.create",
             chat_wrapper(
                 tracer,
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -133,7 +127,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             "ChatCompletion.acreate",
             achat_wrapper(
                 tracer,
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -146,7 +140,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             "Embedding.create",
             embeddings_wrapper(
                 tracer,
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,
@@ -157,7 +151,7 @@ class OpenAIV0Instrumentor(BaseInstrumentor):
             "Embedding.acreate",
             aembeddings_wrapper(
                 tracer,
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/opentelemetry/instrumentation/openai/v1/__init__.py

@@ -49,20 +49,20 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
         meter = get_meter(__name__, __version__, meter_provider)
 
         if is_metrics_enabled():
-            chat_token_counter = meter.create_counter(
-                name="llm.openai.chat_completions.tokens",
+            tokens_histogram = meter.create_histogram(
+                name="gen_ai.client.token.usage",
                 unit="token",
                 description="Number of tokens used in prompt and completions",
             )
 
             chat_choice_counter = meter.create_counter(
-                name="llm.openai.chat_completions.choices",
+                name="gen_ai.client.generation.choices",
                 unit="choice",
                 description="Number of choices returned by chat completions call",
             )
 
             chat_duration_histogram = meter.create_histogram(
-                name="llm.openai.chat_completions.duration",
+                name="gen_ai.client.operation.duration",
                 unit="s",
                 description="Duration of chat completion operation",
             )
@@ -85,7 +85,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             )
         else:
             (
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -98,7 +98,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             "Completions.create",
             chat_wrapper(
                 tracer,
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -114,12 +114,6 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
         )
 
         if is_metrics_enabled():
-            embeddings_token_counter = meter.create_counter(
-                name="llm.openai.embeddings.tokens",
-                unit="token",
-                description="Number of tokens used in prompt and completions",
-            )
-
             embeddings_vector_size_counter = meter.create_counter(
                 name="llm.openai.embeddings.vector_size",
                 unit="element",
@@ -139,7 +133,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             )
         else:
             (
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,
@@ -150,7 +144,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             "Embeddings.create",
             embeddings_wrapper(
                 tracer,
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,
@@ -162,7 +156,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             "AsyncCompletions.create",
             achat_wrapper(
                 tracer,
-                chat_token_counter,
+                tokens_histogram,
                 chat_choice_counter,
                 chat_duration_histogram,
                 chat_exception_counter,
@@ -180,7 +174,7 @@ class OpenAIV1Instrumentor(BaseInstrumentor):
             "AsyncEmbeddings.create",
             aembeddings_wrapper(
                 tracer,
-                embeddings_token_counter,
+                tokens_histogram,
                 embeddings_vector_size_counter,
                 embeddings_duration_histogram,
                 embeddings_exception_counter,

opentelemetry_instrumentation_openai-0.19.0/opentelemetry/instrumentation/openai/version.py

@@ -0,0 +1 @@
+__version__ = "0.19.0"

{opentelemetry_instrumentation_openai-0.18.1 → opentelemetry_instrumentation_openai-0.19.0}/pyproject.toml

@@ -8,7 +8,7 @@ show_missing = true
 
 [tool.poetry]
 name = "opentelemetry-instrumentation-openai"
-version = "0.18.1"
+version = "0.19.0"
 description = "OpenTelemetry OpenAI instrumentation"
 authors = [
   "Gal Kleinman <gal@traceloop.com>",

opentelemetry_instrumentation_openai-0.18.1/opentelemetry/instrumentation/openai/version.py

@@ -1 +0,0 @@
-__version__ = "0.18.1"