feat(core): Add support for tracking provider metadata

packages/core/src/utils/vercel-ai-attributes.ts

@@ -1,3 +1,4 @@
+/* eslint-disable max-lines */
 /**
  * AI SDK Telemetry Attributes
  * Based on https://ai-sdk.dev/docs/ai-sdk-core/telemetry#collected-data
@@ -269,6 +270,15 @@ export const AI_MODEL_PROVIDER_ATTRIBUTE = 'ai.model.provider';
  */
 export const AI_REQUEST_HEADERS_ATTRIBUTE = 'ai.request.headers';
 
+/**
+ * Basic LLM span information
+ * Multiple spans
+ *
+ * Provider specific metadata returned with the generation response
+ * @see https://ai-sdk.dev/docs/ai-sdk-core/telemetry#basic-llm-span-information
+ */
+export const AI_RESPONSE_PROVIDER_METADATA_ATTRIBUTE = 'ai.response.providerMetadata';
+
 /**
  * Basic LLM span information
  * Multiple spans
@@ -792,3 +802,225 @@ export const AI_TOOL_CALL_SPAN_ATTRIBUTES = {
   AI_TELEMETRY_FUNCTION_ID: AI_TELEMETRY_FUNCTION_ID_ATTRIBUTE,
   AI_TELEMETRY_METADATA: AI_TELEMETRY_METADATA_ATTRIBUTE,
 } as const;
+
+// =============================================================================
+// PROVIDER METADATA
+// =============================================================================
+
+/**
+ * OpenAI Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai
+ * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/openai/src/openai-chat-language-model.ts#L397-L416
+ * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/openai/src/responses/openai-responses-language-model.ts#L377C7-L384
+ */
+interface OpenAiProviderMetadata {
+  /**
+   * The number of predicted output tokens that were accepted.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai#predicted-outputs
+   */
+  acceptedPredictionTokens?: number;
+
+  /**
+   * The number of predicted output tokens that were rejected.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai#predicted-outputs
+   */
+  rejectedPredictionTokens?: number;
+
+  /**
+   * The number of reasoning tokens that the model generated.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai#responses-models
+   */
+  reasoningTokens?: number;
+
+  /**
+   * The number of prompt tokens that were a cache hit.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai#responses-models
+   */
+  cachedPromptTokens?: number;
+
+  /**
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/openai#responses-models
+   *
+   * The ID of the response. Can be used to continue a conversation.
+   */
+  responseId?: string;
+}
+
+/**
+ * Anthropic Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/anthropic
+ * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/anthropic/src/anthropic-messages-language-model.ts#L346-L352
+ */
+interface AnthropicProviderMetadata {
+  /**
+   * The number of tokens that were used to create the cache.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/anthropic#cache-control
+   */
+  cacheCreationInputTokens?: number;
+
+  /**
+   * The number of tokens that were read from the cache.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/anthropic#cache-control
+   */
+  cacheReadInputTokens?: number;
+}
+
+/**
+ * Amazon Bedrock Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/amazon-bedrock
+ * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/amazon-bedrock/src/bedrock-chat-language-model.ts#L263-L280
+ */
+interface AmazonBedrockProviderMetadata {
+  /**
+   * @see https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_ConverseTrace.html
+   */
+  trace?: {
+    /**
+     * The guardrail trace object.
+     * @see https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_GuardrailTraceAssessment.html
+     *
+     * This was purposely left as unknown as it's a complex object. This can be typed in the future
+     * if the SDK decides to support bedrock in a more advanced way.
+     */
+    guardrail?: unknown;
+    /**
+     * The request's prompt router.
+     * @see https://docs.aws.amazon.com/bedrock/latest/APIReference/API_runtime_PromptRouterTrace.html
+     */
+    promptRouter?: {
+      /**
+       * The ID of the invoked model.
+       */
+      invokedModelId?: string;
+    };
+  };
+  usage?: {
+    /**
+     * The number of tokens that were read from the cache.
+     * @see https://ai-sdk.dev/providers/ai-sdk-providers/amazon-bedrock#cache-points
+     */
+    cacheReadInputTokens?: number;
+
+    /**
+     * The number of tokens that were written to the cache.
+     * @see https://ai-sdk.dev/providers/ai-sdk-providers/amazon-bedrock#cache-points
+     */
+    cacheWriteInputTokens?: number;
+  };
+}
+
+/**
+ * Google Generative AI Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/google-generative-ai
+ */
+export interface GoogleGenerativeAIProviderMetadata {
+  /**
+   * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/google/src/google-generative-ai-prompt.ts#L28-L30
+   */
+  groundingMetadata: null | {
+    /**
+     * Array of search queries used to retrieve information
+     * @example ["What's the weather in Chicago this weekend?"]
+     *
+     * @see https://ai-sdk.dev/providers/ai-sdk-providers/google-generative-ai#search-grounding
+     */
+    webSearchQueries: string[] | null;
+    /**
+     * Contains the main search result content used as an entry point
+     * The `renderedContent` field contains the formatted content
+     * @see https://ai-sdk.dev/providers/ai-sdk-providers/google-generative-ai#search-grounding
+     */
+    searchEntryPoint?: {
+      renderedContent: string;
+    } | null;
+    /**
+     * Contains details about how specific response parts are supported by search results
+     * @see https://ai-sdk.dev/providers/ai-sdk-providers/google-generative-ai#search-grounding
+     */
+    groundingSupports: Array<{
+      /**
+       * Information about the grounded text segment.
+       */
+      segment: {
+        /**
+         * The start index of the text segment.
+         */
+        startIndex?: number | null;
+        /**
+         * The end index of the text segment.
+         */
+        endIndex?: number | null;
+        /**
+         * The actual text segment.
+         */
+        text?: string | null;
+      };
+      /**
+       * References to supporting search result chunks.
+       */
+      groundingChunkIndices?: number[] | null;
+      /**
+       * Confidence scores (0-1) for each supporting chunk.
+       */
+      confidenceScores?: number[] | null;
+    }> | null;
+  };
+  /**
+   * @see https://github.com/vercel/ai/blob/65e042afde6aad4da9d7a62526ece839eb34f9a5/packages/google/src/google-generative-ai-language-model.ts#L620-L627
+   * @see https://cloud.google.com/vertex-ai/generative-ai/docs/multimodal/configure-safety-filters
+   */
+  safetyRatings?: null | unknown;
+}
+
+/**
+ * DeepSeek Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/deepseek
+ */
+interface DeepSeekProviderMetadata {
+  /**
+   * The number of tokens that were cache hits.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/deepseek#cache-token-usage
+   */
+  promptCacheHitTokens?: number;
+
+  /**
+   * The number of tokens that were cache misses.
+   * @see https://ai-sdk.dev/providers/ai-sdk-providers/deepseek#cache-token-usage
+   */
+  promptCacheMissTokens?: number;
+}
+
+/**
+ * Perplexity Provider Metadata
+ * @see https://ai-sdk.dev/providers/ai-sdk-providers/perplexity
+ */
+interface PerplexityProviderMetadata {
+  /**
+   * Object containing citationTokens and numSearchQueries metrics
+   */
+  usage?: {
+    citationTokens?: number;
+    numSearchQueries?: number;
+  };
+  /**
+   * Array of image URLs when return_images is enabled.
+   *
+   * You can enable image responses by setting return_images: true in the provider options.
+   * This feature is only available to Perplexity Tier-2 users and above.
+   */
+  images?: Array<{
+    imageUrl?: string;
+    originUrl?: string;
+    height?: number;
+    width?: number;
+  }>;
+}
+
+export interface ProviderMetadata {
+  openai?: OpenAiProviderMetadata;
+  anthropic?: AnthropicProviderMetadata;
+  bedrock?: AmazonBedrockProviderMetadata;
+  google?: GoogleGenerativeAIProviderMetadata;
+  deepseek?: DeepSeekProviderMetadata;
+  perplexity?: PerplexityProviderMetadata;
+}

packages/core/src/utils/vercel-ai.ts

@@ -1,14 +1,16 @@
 import type { Client } from '../client';
 import { SEMANTIC_ATTRIBUTE_SENTRY_OP, SEMANTIC_ATTRIBUTE_SENTRY_ORIGIN } from '../semanticAttributes';
 import type { Event } from '../types-hoist/event';
-import type { Span, SpanAttributes, SpanJSON, SpanOrigin } from '../types-hoist/span';
+import type { Span, SpanAttributes, SpanAttributeValue, SpanJSON, SpanOrigin } from '../types-hoist/span';
 import { spanToJSON } from './spanUtils';
+import type { ProviderMetadata } from './vercel-ai-attributes';
 import {
   AI_MODEL_ID_ATTRIBUTE,
   AI_MODEL_PROVIDER_ATTRIBUTE,
   AI_PROMPT_ATTRIBUTE,
   AI_PROMPT_MESSAGES_ATTRIBUTE,
   AI_PROMPT_TOOLS_ATTRIBUTE,
+  AI_RESPONSE_PROVIDER_METADATA_ATTRIBUTE,
   AI_RESPONSE_TEXT_ATTRIBUTE,
   AI_RESPONSE_TOOL_CALLS_ATTRIBUTE,
   AI_TELEMETRY_FUNCTION_ID_ATTRIBUTE,
@@ -96,6 +98,8 @@ function processEndedVercelAiSpan(span: SpanJSON): void {
   renameAttributeKey(attributes, AI_TOOL_CALL_ARGS_ATTRIBUTE, 'gen_ai.tool.input');
   renameAttributeKey(attributes, AI_TOOL_CALL_RESULT_ATTRIBUTE, 'gen_ai.tool.output');
 
+  addProviderMetadataToAttributes(attributes);
+
   // Change attributes namespaced with `ai.X` to `vercel.ai.X`
   for (const key of Object.keys(attributes)) {
     if (key.startsWith('ai.')) {
@@ -234,3 +238,85 @@ export function addVercelAiProcessors(client: Client): void {
   // Note: We cannot do this on `spanEnd`, because the span cannot be mutated anymore at this point
   client.addEventProcessor(Object.assign(vercelAiEventProcessor, { id: 'VercelAiEventProcessor' }));
 }
+
+function addProviderMetadataToAttributes(attributes: SpanAttributes): void {
+  const providerMetadata = attributes[AI_RESPONSE_PROVIDER_METADATA_ATTRIBUTE] as string | undefined;
+  if (providerMetadata) {
+    try {
+      const providerMetadataObject = JSON.parse(providerMetadata) as ProviderMetadata;
+      if (providerMetadataObject.openai) {
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cached',
+          providerMetadataObject.openai.cachedPromptTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.output_tokens.reasoning',
+          providerMetadataObject.openai.reasoningTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.output_tokens.prediction_accepted',
+          providerMetadataObject.openai.acceptedPredictionTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.output_tokens.prediction_rejected',
+          providerMetadataObject.openai.rejectedPredictionTokens,
+        );
+        setAttributeIfDefined(attributes, 'gen_ai.conversation.id', providerMetadataObject.openai.responseId);
+      }
+
+      if (providerMetadataObject.anthropic) {
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cached',
+          providerMetadataObject.anthropic.cacheReadInputTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cache_write',
+          providerMetadataObject.anthropic.cacheCreationInputTokens,
+        );
+      }
+
+      if (providerMetadataObject.bedrock?.usage) {
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cached',
+          providerMetadataObject.bedrock.usage.cacheReadInputTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cache_write',
+          providerMetadataObject.bedrock.usage.cacheWriteInputTokens,
+        );
+      }
+
+      if (providerMetadataObject.deepseek) {
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cached',
+          providerMetadataObject.deepseek.promptCacheHitTokens,
+        );
+        setAttributeIfDefined(
+          attributes,
+          'gen_ai.usage.input_tokens.cache_miss',
+          providerMetadataObject.deepseek.promptCacheMissTokens,
+        );
+      }
+    } catch {
+      // Ignore
+    }
+  }
+}
+
+/**
+ * Sets an attribute only if the value is not null or undefined.
+ */
+function setAttributeIfDefined(attributes: SpanAttributes, key: string, value: SpanAttributeValue | undefined): void {
+  if (value != null) {
+    attributes[key] = value;
+  }
+}