fix(ai-sdk): report cached tokens usage (#839)

Author: galkleinman

packages/ai-semantic-conventions/src/SemanticAttributes.ts

@@ -35,6 +35,10 @@ export const SpanAttributes = {
   LLM_USAGE_COMPLETION_TOKENS: "gen_ai.usage.completion_tokens",
   LLM_USAGE_INPUT_TOKENS: ATTR_GEN_AI_USAGE_INPUT_TOKENS,
   LLM_USAGE_OUTPUT_TOKENS: ATTR_GEN_AI_USAGE_OUTPUT_TOKENS,
+  GEN_AI_USAGE_CACHE_CREATION_INPUT_TOKENS:
+    "gen_ai.usage.cache_creation_input_tokens",
+  GEN_AI_USAGE_CACHE_READ_INPUT_TOKENS: "gen_ai.usage.cache_read_input_tokens",
+  GEN_AI_USAGE_REASONING_TOKENS: "gen_ai.usage.reasoning_tokens",
 
   GEN_AI_AGENT_NAME: "gen_ai.agent.name",
 

packages/instrumentation-anthropic/package.json

@@ -46,7 +46,7 @@
     "tslib": "^2.8.1"
   },
   "devDependencies": {
-    "@anthropic-ai/sdk": "^0.56.0",
+    "@anthropic-ai/sdk": "^0.71.0",
     "@opentelemetry/context-async-hooks": "^2.0.1",
     "@opentelemetry/sdk-trace-node": "^2.0.1",
     "@pollyjs/adapter-fetch": "^6.0.7",

packages/instrumentation-anthropic/recordings/Test-Anthropic-instrumentation_3769946143/should-set-attributes-in-span-for-completions-streaming_2198009633/recording.har

@@ -317,8 +317,9 @@ export class AnthropicInstrumentation extends InstrumentationBase {
             usage: {
               input_tokens: 0,
               output_tokens: 0,
-              cache_creation_input_tokens: 0,
-              cache_read_input_tokens: 0,
+              cache_creation: null,
+              cache_creation_input_tokens: null,
+              cache_read_input_tokens: null,
               server_tool_use: null,
               service_tier: null,
             },

packages/instrumentation-anthropic/recordings/Test-Anthropic-instrumentation_3769946143/should-set-attributes-in-span-for-completions_1224394582/recording.har

@@ -91,100 +91,6 @@ describe("Test Anthropic instrumentation", async function () {
     context.disable();
   });
 
-  it("should set attributes in span for completions", async () => {
-    const result = await anthropic.completions.create({
-      model: "claude-2",
-      max_tokens_to_sample: 300,
-      prompt: `${AnthropicModule.HUMAN_PROMPT} Tell me a joke about OpenTelemetry${AnthropicModule.AI_PROMPT}`,
-    });
-
-    const spans = memoryExporter.getFinishedSpans();
-    const completionSpan = spans.find(
-      (span) => span.name === "anthropic.completion",
-    );
-
-    assert.ok(result);
-    assert.ok(completionSpan);
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_REQUEST_MODEL}`],
-      "claude-2",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_REQUEST_MAX_TOKENS}`],
-      300,
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_RESPONSE_MODEL}`],
-      "claude-2.1",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_PROMPTS}.0.role`],
-      "user",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_PROMPTS}.0.content`],
-      `${AnthropicModule.HUMAN_PROMPT} Tell me a joke about OpenTelemetry${AnthropicModule.AI_PROMPT}`,
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_COMPLETIONS}.0.role`],
-      "assistant",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_COMPLETIONS}.0.content`],
-      result.completion,
-    );
-  }).timeout(30000);
-
-  it("should set attributes in span for completions (streaming)", async () => {
-    const result = await anthropic.completions.create({
-      model: "claude-2",
-      max_tokens_to_sample: 300,
-      prompt: `${AnthropicModule.HUMAN_PROMPT} Tell me a joke about OpenTelemetry${AnthropicModule.AI_PROMPT}`,
-      stream: true,
-    });
-
-    let completion = "";
-    for await (const chunk of result) {
-      assert.ok(chunk);
-      completion += chunk.completion;
-    }
-
-    const spans = memoryExporter.getFinishedSpans();
-    const completionSpan = spans.find(
-      (span) => span.name === "anthropic.completion",
-    );
-
-    assert.ok(completionSpan);
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_REQUEST_MODEL}`],
-      "claude-2",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_REQUEST_MAX_TOKENS}`],
-      300,
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_RESPONSE_MODEL}`],
-      "claude-2.1",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_PROMPTS}.0.role`],
-      "user",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_PROMPTS}.0.content`],
-      `${AnthropicModule.HUMAN_PROMPT} Tell me a joke about OpenTelemetry${AnthropicModule.AI_PROMPT}`,
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_COMPLETIONS}.0.role`],
-      "assistant",
-    );
-    assert.strictEqual(
-      completionSpan.attributes[`${SpanAttributes.LLM_COMPLETIONS}.0.content`],
-      completion,
-    );
-  }).timeout(30000);
-
   it("should set attributes in span for messages", async () => {
     const message = await anthropic.messages.create({
       max_tokens: 1024,

packages/instrumentation-anthropic/src/instrumentation.ts

@@ -93,9 +93,10 @@
   "gitHead": "ef1e70d6037f7b5c061056ef2be16e3f55f02ed5",
   "devDependencies": {
     "@ai-sdk/amazon-bedrock": "^3.0.10",
+    "@ai-sdk/anthropic": "^2.0.53",
     "@ai-sdk/google": "^2.0.8",
     "@ai-sdk/openai": "^2.0.19",
-    "@anthropic-ai/sdk": "^0.56.0",
+    "@anthropic-ai/sdk": "^0.71.0",
     "@aws-sdk/client-bedrock-runtime": "^3.848.0",
     "@google-cloud/aiplatform": "^4.4.0",
     "@google-cloud/vertexai": "^1.10.0",