[BUG]: no token count on langgraph instrumentation

[mgathena]

Where do you use Phoenix

Self-hosted

What version of Phoenix are you using?

12.14.2

What happened?

total tokens (and hence cost) is 0

the relevant snippet

exporter = OTLPSpanExporter(
    endpoint=f"{os.getenv('PHOENIX_COLLECTOR_ENDPOINT')}/v1/traces",
    headers={"Authorization": f"Bearer {os.getenv('PHOENIX_API_KEY')}"},
)
tracer_provider.add_span_processor(SimpleSpanProcessor(exporter))

LangChainInstrumentor().instrument(tracer_provider=tracer_provider)

I even tried adding OpenAIInstrumentor but that too does not log token count etc. My model is gpt-4.1 (which maps to gpt-4.1-2025-04-14)

Additional information

openinference-instrumentation-langchain==0.1.55

[dosubot]

There are a few known reasons why token count and cost can show as zero in Phoenix when using LangChainInstrumentor and OpenAIInstrumentor, especially with langgraph workflows and recent OpenAI models:

• Token usage is only reported if the upstream LLM API (like OpenAI's gpt-4.1) returns usage data. If the API response doesn't include token counts (which is common for streaming responses unless you explicitly request usage with something like stream_options={"include_usage": True}), the instrumentation can't populate these fields, and Phoenix will show zero tokens and cost. There is no fallback logic to estimate tokens if the provider doesn't supply them details streaming caveat.

• In complex workflows like those built with langgraph, cost and token aggregation in Phoenix relies on direct parent-child span relationships. If your workflow's root span doesn't directly parent all cost-bearing spans, cumulative costs and tokens may not appear at the root or in summary views. This is a structural limitation in the current OpenInference cost propagation logic see explanation.

• Using multiple instrumentors (e.g., LangChain and OpenAI) at once can cause nested/conflicting spans, which may result in missing or zero token counts. It's best to use only one instrumentor for LLM calls best practice.

• There is a known bug where token counts display as zero in the Phoenix UI when using the LangChain auto-instrumentor, even though the correct counts are output elsewhere. This has been observed with other providers (like Vertex) and may also affect OpenAI bug report.

To troubleshoot: check if your LLM client is requesting usage data for streaming calls, inspect the raw span payloads for token usage attributes, and ensure only one instrumentor is active. For langgraph, be aware that root-level aggregation may not reflect all descendant costs/tokens due to current limitations.

If you need more granular or reliable token counts, you may need to add explicit token counting logic in your pipeline or monitor for updates in OpenInference and Phoenix addressing these limitations.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[cephalization]

Moving this to OpenInference repository for triage

[dosubot]

This is a known limitation with the current LangChain instrumentation: token counts can be zero if the instrumentor doesn't recognize the nested usage keys returned by newer OpenAI models like gpt-4.1. The LangChainInstrumentor in openinference-instrumentation-langchain v0.1.55 doesn't fully extract token counts from nested fields such as input_tokens_details.cached_tokens or similar structures in the usage metadata, which are present in gpt-4.1 responses. OpenAIInstrumentor handles these nested keys, but if your LangChain workflow wraps the OpenAI client, the direct OpenAI instrumentation may not see the raw responses and thus also miss the counts (source).

A recent update to the LangChainInstrumentor (after v0.1.55) improves token counting by parsing these nested usage metadata fields, so upgrading to the latest version of openinference-instrumentation-langchain and ensuring your langchain_core is at least v0.3.9 should resolve the issue (source). If you can't upgrade, you may be able to work around this by instrumenting the OpenAI client directly (outside of LangChain), but for full support, updating the instrumentation is recommended.

Also, make sure your tracer provider is registered with auto_instrument=True and batch=True if using the latest Phoenix integration patterns, as this can affect how spans and attributes are captured.

If you are using streaming responses, note that OpenAI APIs only return usage data if you explicitly request it (e.g., with stream_options={'include_usage': True}), and some wrappers may not support this yet (source).

In summary: upgrade openinference-instrumentation-langchain and langchain_core, and check your instrumentation setup for compatibility with gpt-4.1 usage metadata.

To reply, just mention @dosu.

---

^{How did I do? Good | Irrelevant | Incorrect | Verbose | Hallucination | Report 🐛 | Other}    

[mgathena]

Adding stream_options={'include_usage': True}) works.