fix: stop overwriting multi-message INPUT_VALUE in OpenAI instrumentor

python/frameworks/openai/tests/test_span_io_handler.py

@@ -0,0 +1,86 @@
+"""
+Regression test for issue #151: multi-turn conversation input was being
+dropped from the span. _process_input_data set INPUT_VALUE three times
+on the same span, and the last two writes overwrote the first, correct
+one, leaving only the first message's text on the span.
+"""
+
+import json
+
+import pytest
+from opentelemetry.sdk.trace import TracerProvider
+from opentelemetry.sdk.trace.export import SimpleSpanProcessor
+from opentelemetry.sdk.trace.export.in_memory_span_exporter import (
+    InMemorySpanExporter,
+)
+
+from fi_instrumentation.fi_types import SpanAttributes
+from traceai_openai._span_io_handler import _process_input_data
+from traceai_openai._with_span import _WithSpan
+
+
+@pytest.fixture
+def recorder():
+    exporter = InMemorySpanExporter()
+    provider = TracerProvider()
+    provider.add_span_processor(SimpleSpanProcessor(exporter))
+    tracer = provider.get_tracer(__name__)
+    return tracer, exporter
+
+
+def _get_input_value(exporter):
+    spans = exporter.get_finished_spans()
+    assert len(spans) == 1
+    return spans[0].attributes.get(SpanAttributes.INPUT_VALUE)
+
+
+def test_multi_message_input_is_not_dropped(recorder):
+    tracer, exporter = recorder
+    messages = [
+        {"role": "system", "content": "You are a helpful assistant."},
+        {"role": "user", "content": "What is the capital of France?"},
+        {"role": "assistant", "content": "Paris."},
+        {"role": "user", "content": "And Germany?"},
+    ]
+
+    with tracer.start_as_current_span("openai.chat") as span:
+        _process_input_data(messages, _WithSpan(span))
+
+    input_value = _get_input_value(exporter)
+    assert input_value is not None
+
+    parsed = json.loads(input_value)
+    assert len(parsed) == len(messages), (
+        "Expected all messages to survive in INPUT_VALUE, but only "
+        f"{len(parsed)} of {len(messages)} were present."
+    )
+    assert [m["content"] for m in parsed] == [m["content"] for m in messages]
+
+
+def test_multimodal_message_list_preserves_all_messages(recorder):
+    tracer, exporter = recorder
+    messages = [
+        {"role": "user", "content": "Describe this image."},
+        {
+            "role": "user",
+            "content": [
+                {"type": "text", "text": "What is in this picture?"},
+                {
+                    "type": "image_url",
+                    "image_url": {"url": "https://example.com/cat.png"},
+                },
+            ],
+        },
+    ]
+
+    with tracer.start_as_current_span("openai.chat") as span:
+        _process_input_data(messages, _WithSpan(span))
+
+    input_value = _get_input_value(exporter)
+    parsed = json.loads(input_value)
+    assert len(parsed) == 2
+
+    spans = exporter.get_finished_spans()
+    images_value = spans[0].attributes.get(SpanAttributes.INPUT_IMAGES)
+    assert images_value is not None
+    assert json.loads(images_value) == ["https://example.com/cat.png"]

python/frameworks/openai/traceai_openai/_span_io_handler.py

@@ -19,7 +19,6 @@ def _process_input_data(input_data: Any, span: _WithSpan) -> None:
     if isinstance(input_data, list):
         input_content = []
         input_images = []
-        eval_input = []
         for msg in input_data:
             if isinstance(msg, dict):
                 msg_content = msg.get("content", "")
@@ -29,18 +28,11 @@ def _process_input_data(input_data: Any, span: _WithSpan) -> None:
                         if isinstance(item, dict):
                             if item.get("type") == "text":
                                 filtered_content.append(item)
-                                eval_input.append(item.get("text", ""))
                             elif item.get("type") == "image_url":
                                 url_data = item.get("image_url", {})
                                 url = url_data.get("url")
                                 if url:
                                     input_images.append(url)
-                                    image_index = len(input_images) - 1
-                                    eval_input.append(
-                                        "{{"
-                                        + f"{SpanAttributes.INPUT_IMAGES}.{image_index}"
-                                        + "}}"
-                                    )
                     if filtered_content:
                         msg_dict = msg.copy()
                         msg_dict["content"] = filtered_content
@@ -49,18 +41,15 @@ def _process_input_data(input_data: Any, span: _WithSpan) -> None:
                         continue
                 else:
                     input_content.append(msg)
-                    eval_input.append(msg_content)
+        # INPUT_VALUE holds the full multi-message conversation as JSON.
+        # Do not overwrite it later with a single-message string; that was
+        # the cause of multi-turn context being dropped from the span (#151).
         if input_content:
             input_value = json.dumps(input_content, ensure_ascii=False)
             span.set_attribute(SpanAttributes.INPUT_VALUE, input_value)
         if input_images:
             images_value = json.dumps(input_images, ensure_ascii=False)
             span.set_attribute(SpanAttributes.INPUT_IMAGES, images_value)
-        if eval_input:
-            eval_input_str = " \n ".join(map(str, eval_input))
-            span.set_attribute(SpanAttributes.INPUT_VALUE, eval_input_str)
-        if eval_input and len(eval_input) > 0:
-            span.set_attribute(SpanAttributes.INPUT_VALUE, eval_input[0])
     else:
         try:
             input_str = json.dumps(input_data, ensure_ascii=False).strip()