refactor(otel): trim the message cap docstrings to one line each

Co-Authored-By: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

litellm/integrations/otel/mappers/openinference.py

@@ -104,22 +104,13 @@ class OpenInferenceMapper:
 
     @staticmethod
     def _indexed_split(inputs: int, outputs: int) -> tuple[int, int]:
-        """How many prompt and response messages get per-index attributes.
-
-        Both directions share one span-wide allowance. The response is reserved at
-        least half of it, so a long prompt can never push the completion off the
-        span, and the prompt takes whatever the response leaves unused.
-        """
+        """Prompt and response share one allowance; the response is reserved at least half of it."""
         indexed_out: Final = min(outputs, max(_MAX_INDEXED_MESSAGES // 2, _MAX_INDEXED_MESSAGES - inputs))
         return _MAX_INDEXED_MESSAGES - indexed_out, indexed_out
 
     @staticmethod
     def _prompt_positions(total: int, indexed: int) -> tuple[int, ...]:
-        """Which prompt messages get per-index attributes: message 0 and the most recent turns.
-
-        A value length limit clips the ``input.value`` blob, so the system prompt and the
-        live turn each keep a short key of their own. The middle of a long prompt does not.
-        """
+        """Prompt messages that get per-index attributes: message 0 and the most recent turns."""
         if total <= indexed:
             return tuple(range(total))
         return (0, *range(total - indexed + 1, total))

litellm/integrations/otel/mappers/utils.py

@@ -33,17 +33,10 @@ core telemetry no matter how many vocabularies are configured.
 
 
 MAX_MESSAGE_ATTRS_PER_SPAN: Final = DEFAULT_SPAN_ATTRIBUTE_LIMIT // 8
-"""Span-wide ceiling on attributes spent spelling out chat messages per index.
+"""Span-wide ceiling on per-index chat message attributes, prompt and response together.
 
-A conversation is the other unbounded family: two attributes per message, for
-the prompt and the response alike, on the same span. Past a few dozen turns the
-family alone exceeds the span attribute limit and evicts the core telemetry
-written before it. The ceiling covers both directions together, since a budget
-handed to each direction separately doubles. An eighth is the largest share
-that still fits beside the tool ceiling and the core of every vocabulary at
-once, request parameters, cost breakdown and identity included. The complete
-conversation still rides the JSON blob attributes; only the per-index
-convenience keys are capped.
+An eighth is the largest share that still fits beside the tool ceiling and the core
+of every vocabulary at once. The complete conversation still rides the JSON blobs.
 """
 
 

tests/test_litellm/integrations/otel/test_otel_v2_emitter.py

@@ -485,13 +485,7 @@ def _indexed_message_count(attributes, prefix):
 
 @pytest.mark.parametrize("turns", [60, 200])
 def test_long_conversation_does_not_evict_core_attributes(turns):
-    """Per-message OpenInference attributes must never crowd core telemetry off the span.
-
-    With content capture on, the OpenInference vocabulary spells every prompt and
-    response message out as two per-index attributes. A few dozen turns overruns
-    the OTel SDK's 128-attribute span limit, which evicts oldest-first, so the
-    ``gen_ai.*`` set written before it is what disappears.
-    """
+    """Per-message OpenInference attributes must never crowd core telemetry off the span."""
     span = _conversation_span(["genai", "openinference"], _conversation_payload(turns))
     a = span.attributes
 
@@ -522,12 +516,7 @@ def test_short_conversation_keeps_every_message_indexed():
 
 
 def test_indexed_prompt_keeps_opener_and_latest_turns_under_a_value_length_limit(monkeypatch):
-    """The per-index keys are the only untruncated copy once the SDK clips string values.
-
-    Operators bound attribute sizes with ``OTEL_SPAN_ATTRIBUTE_VALUE_LENGTH_LIMIT``,
-    which cuts the ``input.value`` blob short. The system prompt and the live turn
-    then have to survive as their own short keys, whatever the conversation length.
-    """
+    """The system prompt and the live turn keep their own keys once the SDK clips ``input.value``."""
     monkeypatch.setenv("OTEL_SPAN_ATTRIBUTE_VALUE_LENGTH_LIMIT", "256")
     payload = _conversation_payload(60)
     payload["messages"][0] = {"role": "system", "content": "be terse"}
@@ -547,11 +536,7 @@ def test_indexed_prompt_keeps_opener_and_latest_turns_under_a_value_length_limit
 
 
 def test_message_cap_is_shared_across_input_and_output():
-    """One span-wide allowance covers both directions, and the response always keeps a share.
-
-    A long prompt takes what a single reply leaves over, and a many-choice reply
-    cannot take the whole allowance away from the prompt either.
-    """
+    """One span-wide allowance covers both directions, and the response always keeps a share."""
     long_prompt = _conversation_span(["genai", "openinference"], _conversation_payload(60, choices=1)).attributes
     many_choices = _conversation_span(["genai", "openinference"], _conversation_payload(60, choices=20)).attributes
 
@@ -569,13 +554,7 @@ def test_message_cap_is_shared_across_input_and_output():
 
 
 def test_fully_populated_span_with_every_vocabulary_stays_within_the_attribute_limit():
-    """Every capped family maxed at once still leaves the whole core intact.
-
-    Every vocabulary in the registry plus ``legacy``, every request parameter,
-    every cost component, a hundred-plus tools, a two-hundred-turn prompt and
-    twenty choices is the worst case the two span-wide ceilings have to absorb
-    together.
-    """
+    """Every capped family maxed at once still leaves the whole core intact."""
     payload = _conversation_payload(
         200,
         choices=20,