Add support for OpenAI llmOutput.tokenUsage to support prompt_tokens_details and completion_tokens_details (prompt caching and reasoning tokens)

[a-i-joe]

Checked

• I searched existing ideas and did not find a similar one
• I added a very descriptive title
• I've clearly described the feature request and motivation for it

Feature request

Here's the pertinent code which selectively maps fields from the OpenAI response usage field

langchainjs/libs/langchain-openai/src/llms.ts

Line 343 in 8e12ef9

const {

Currently it only map out three fields:

const {
        completion_tokens: completionTokens,
        prompt_tokens: promptTokens,
        total_tokens: totalTokens,
      } = data.usage
        ? data.usage
        : {
            completion_tokens: undefined,
            prompt_tokens: undefined,
            total_tokens: undefined,
          };

To be able to track other token details, such as cached tokens, and tokens used for reasoning by reasoning models, more fields must be mapped. Here's the modern openai usage schema:

"usage": {
  "prompt_tokens": 2006,
  "completion_tokens": 300,
  "total_tokens": 2306,
  "prompt_tokens_details": {
    "cached_tokens": 1920
  },
  "completion_tokens_details": {
    "reasoning_tokens": 0,
    "accepted_prediction_tokens": 0,
    "rejected_prediction_tokens": 0
  }
}

Motivation

Cached tokens, and tokens used for reasoning by reasoning models are important cost factors for cost tracking and optimization. Without access to these, there is significant disadvantage to using langchainjs.

Proposal (If applicable)

Expand the TokenUsage interface, and map the additional fields.

I'm planning on forking the code to do this, and I'd be happy to create a PR if desired. I intend to map it to this structure:

interface TokenUsage {
  completionTokens?: number;
  completionTokensDetails?: {
    reasoningTokens?: number;
    acceptedPredictionTokens?: number;
    rejectedPredictionTokens?: number;
  };
  promptTokens?: number;
  promptTokensDetails?: {
    cachedTokens?: number;
  };
  totalTokens?: number;
}

Feedback welcome. Thank you.

[a-i-joe]

See PR #8407