fix(instrumentation-bedrock): gen_ai.usage.total_tokens is NaN (string) for Titan embedding responses

[BrontoStephen]

Summary

For Amazon Titan embedding responses (e.g. amazon.titan-embed-text-v2:0), the Bedrock instrumentation emits gen_ai.usage.total_tokens = NaN. Embedding responses contain inputTextTokenCount but no output-token field, and the total is computed as input + undefined, which in JS silently evaluates to NaN. The attribute is then serialized as the string "NaN" on the exported span.

Environment

• @traceloop/instrumentation-bedrock via @traceloop/node-server-sdk
• Node.js 22, InvokeModel with amazon.titan-embed-text-v2:0
• Bug is present on main (1faf69d)

Observed span attributes

gen_ai.operation.name:      embeddings
gen_ai.request.model:       titan-embed-text-v2:0
gen_ai.usage.input_tokens:  117
gen_ai.usage.total_tokens:  "NaN"   <-- string, poisons the attribute
(gen_ai.usage.output_tokens absent)

In our production telemetry, 100% of spans carrying total_tokens from this code path have the value "NaN" (152 spans over 7 days).

Root cause

Two sites guard only the input side before adding input + output:

1. Non-streaming Titan path — titanOutputTokens comes from response["results"]?.[0]?.["tokenCount"], which doesn't exist on embedding responses:
   

   openllmetry-js/packages/instrumentation-bedrock/src/instrumentation.ts

   Lines 642 to 648 in 1faf69d

   	...(titanInputTokens != null
   	? {
   	[ATTR_GEN_AI_USAGE_INPUT_TOKENS]: titanInputTokens,
   	[ATTR_GEN_AI_USAGE_OUTPUT_TOKENS]: titanOutputTokens,
   	[SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS]:
   	titanInputTokens + titanOutputTokens,
   	}

2. Streaming path ("Titan includes token counts on the final chunk") — same pattern with response["totalOutputTextTokenCount"]:
   

   openllmetry-js/packages/instrumentation-bedrock/src/instrumentation.ts

   Lines 603 to 612 in 1faf69d

   	...(response["inputTextTokenCount"] != null
   	? {
   	[ATTR_GEN_AI_USAGE_INPUT_TOKENS]:
   	response["inputTextTokenCount"],
   	[ATTR_GEN_AI_USAGE_OUTPUT_TOKENS]:
   	response["totalOutputTextTokenCount"],
   	[SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS]:
   	response["inputTextTokenCount"] +
   	response["totalOutputTextTokenCount"],
   	}

The existing tests only cover the happy path where both counts exist (amazon.test.ts asserts input + output), so the embedding case is never exercised.

Impact

Beyond the value being wrong, emitting the string "NaN" has a nasty downstream effect: OTel backends that infer attribute types see a string and lock gen_ai.usage.total_tokens to STRING for the dataset, which breaks numeric aggregations (SUM/AVG/percentiles) over the attribute even for spans where the value would have been numeric.

Suggested fix

Only emit GEN_AI_USAGE_TOTAL_TOKENS when both operands are numbers, and treat missing output as 0 for operations that legitimately have no output tokens (embeddings are billed on input only):

...(titanInputTokens != null
  ? {
      [ATTR_GEN_AI_USAGE_INPUT_TOKENS]: titanInputTokens,
      ...(titanOutputTokens != null
        ? { [ATTR_GEN_AI_USAGE_OUTPUT_TOKENS]: titanOutputTokens }
        : {}),
      [SpanAttributes.GEN_AI_USAGE_TOTAL_TOKENS]:
        titanInputTokens + (titanOutputTokens ?? 0),
    }
  : {}),

Same guard for the streaming site. Happy to open a PR if that approach sounds right.