Description

When using LangChain with streaming enabled (streaming=True or stream_options={"include_usage": True}), the gen_ai.usage.input_tokens and gen_ai.usage.output_tokens span attributes are missing from traces. This prevents accurate token usage tracking and cost attribution for streaming LLM calls.

Environment

• splunk-otel-instrumentation-langchain: 0.1.x
• langchain-openai: 0.3.x (or langchain-anthropic)
• Provider: OpenAI, Anthropic, or any provider using streaming with usage_metadata

Steps to Reproduce

from langchain_openai import ChatOpenAI
from langchain_core.messages import HumanMessage

# Create LLM with streaming enabled
llm = ChatOpenAI(
    model="gpt-4o-mini",
    streaming=True,
    model_kwargs={"stream_options": {"include_usage": True}}
)

# Invoke - response.usage_metadata WILL have tokens
response = llm.invoke([HumanMessage(content="Say hello")])
print(response.usage_metadata)  # {'input_tokens': 40, 'output_tokens': 55, ...}

# But the trace span will be missing token attributes

Expected Behavior

Span: chat gpt-4o-mini
Attributes:
  gen_ai.usage.input_tokens: 40
  gen_ai.usage.output_tokens: 55

Actual Behavior

Span: chat gpt-4o-mini
Attributes:
  gen_ai.request.model: gpt-4o-mini
  gen_ai.response.model: gpt-4o-mini-2024-07-18
  # No gen_ai.usage.* attributes

Evidence

Direct comparison from live testing with Cisco/OpenAI endpoint:

{
  "traceId": "49f46b971825fa178f4ae86812a9f1d2",
  "operationName": "chat gpt-4o-mini",
  "tags": {
    "gen_ai.request.model": "gpt-4o-mini",
    "gen_ai.response.model": "gpt-4o-mini-2024-07-18",
    "gen_ai.usage.input_tokens": 31,
    "gen_ai.usage.output_tokens": 5,
    "gen_ai.provider.name": "openai"
  }
}

📊 Streaming Trace (tokens missing) - click to expand

Trace ID: 77432872a967d4321701ce1f22032d8c

{
  "traceId": "77432872a967d4321701ce1f22032d8c",
  "operationName": "chat gpt-4o-mini",
  "tags": {
    "gen_ai.request.model": "gpt-4o-mini",
    "gen_ai.response.model": "gpt-4o-mini-2024-07-18",
    "gen_ai.provider.name": "openai"
  }
}

❌ NO gen_ai.usage.input_tokens or gen_ai.usage.output_tokens!

Yet Python showed tokens were available:

response.usage_metadata: {'input_tokens': 40, 'output_tokens': 55, 'total_tokens': 95}

{
  "traceId": "303595c0d1031acdae9bacd46083d87b",
  "operationName": "chat gpt-4o-mini",
  "tags": {
    "gen_ai.request.model": "gpt-4o-mini",
    "gen_ai.response.model": "gpt-4o-mini-2024-07-18",
    "gen_ai.usage.input_tokens": 40,
    "gen_ai.usage.output_tokens": 57,
    "gen_ai.provider.name": "openai"
  }
}

Full trace evidence:

• Before fix - streaming tokens MISSING
• After fix - streaming tokens PRESENT

Root Cause Analysis

The current code only extracts tokens from llm_output.token_usage:

# Current code - ONLY checks llm_output (non-streaming path)
llm_output = getattr(response, "llm_output", {}) or {}
usage = llm_output.get("usage") or llm_output.get("token_usage") or {}
inv.input_tokens = usage.get("prompt_tokens")
inv.output_tokens = usage.get("completion_tokens")

In streaming mode:

• llm_output is empty or missing token data
• Tokens are in response.generations[0][0].message.usage_metadata instead

Affected Providers

API References

OpenAI:

When stream_options: {"include_usage": true} is set, token usage is returned in the final streaming chunk via usage field.
— OpenAI API Reference

Anthropic:

The token counts shown in the usage field of the message_delta event are cumulative.
— Anthropic Streaming Messages Docs

Impact

• ❌ Cost tracking broken for streaming calls (often 50%+ of production traffic)
• ❌ Token budgets unenforceable - can't track streaming token consumption
• ❌ Billing reconciliation impossible - streaming costs not attributed
• ❌ Performance analysis incomplete - can't correlate tokens with latency for streaming
• ❌ Dashboards show gaps - token metrics missing for streaming spans