Properly populate usage_metadata for Anthropic models

libs/langchain-anthropic/package.json

@@ -35,7 +35,7 @@
   "author": "LangChain",
   "license": "MIT",
   "dependencies": {
-    "@anthropic-ai/sdk": "^0.27.3",
+    "@anthropic-ai/sdk": "^0.32.1",
     "fast-xml-parser": "^4.4.1",
     "zod": "^3.22.4",
     "zod-to-json-schema": "^3.22.4"

libs/langchain-anthropic/src/tests/chat_models.int.test.ts

@@ -27,7 +27,6 @@ test("Test ChatAnthropic", async () => {
   });
   const message = new HumanMessage("Hello!");
   const res = await chat.invoke([message]);
-  // console.log({ res });
   expect(res.response_metadata.usage).toBeDefined();
 });
 
@@ -713,12 +712,12 @@ test("system prompt caching", async () => {
   ];
   const res = await model.invoke(messages);
   expect(
-    res.response_metadata.usage.cache_creation_input_tokens
+    res.usage_metadata?.input_token_details?.cache_creation
   ).toBeGreaterThan(0);
-  expect(res.response_metadata.usage.cache_read_input_tokens).toBe(0);
+  expect(res.usage_metadata?.input_token_details?.cache_read).toBe(0);
   const res2 = await model.invoke(messages);
-  expect(res2.response_metadata.usage.cache_creation_input_tokens).toBe(0);
-  expect(res2.response_metadata.usage.cache_read_input_tokens).toBeGreaterThan(
+  expect(res2.usage_metadata?.input_token_details?.cache_creation).toBe(0);
+  expect(res2.usage_metadata?.input_token_details?.cache_read).toBeGreaterThan(
     0
   );
   const stream = await model.stream(messages);
@@ -727,8 +726,8 @@ test("system prompt caching", async () => {
     agg = agg === undefined ? chunk : concat(agg, chunk);
   }
   expect(agg).toBeDefined();
-  expect(agg!.response_metadata.usage.cache_creation_input_tokens).toBe(0);
-  expect(agg!.response_metadata.usage.cache_read_input_tokens).toBeGreaterThan(
+  expect(agg!.usage_metadata?.input_token_details?.cache_creation).toBe(0);
+  expect(agg!.usage_metadata?.input_token_details?.cache_read).toBeGreaterThan(
     0
   );
 });
@@ -771,12 +770,12 @@ test.skip("tool caching", async () => {
   const res = await model.invoke(messages);
   console.log(res);
   expect(
-    res.response_metadata.usage.cache_creation_input_tokens
+    res.usage_metadata?.input_token_details?.cache_creation
   ).toBeGreaterThan(0);
-  expect(res.response_metadata.usage.cache_read_input_tokens).toBe(0);
+  expect(res.usage_metadata?.input_token_details?.cache_read).toBe(0);
   const res2 = await model.invoke(messages);
-  expect(res2.response_metadata.usage.cache_creation_input_tokens).toBe(0);
-  expect(res2.response_metadata.usage.cache_read_input_tokens).toBeGreaterThan(
+  expect(res2.usage_metadata?.input_token_details?.cache_creation).toBe(0);
+  expect(res2.usage_metadata?.input_token_details?.cache_read).toBeGreaterThan(
     0
   );
 });
@@ -791,7 +790,7 @@ test.skip("Test ChatAnthropic with custom client", async () => {
   const message = new HumanMessage("Hello!");
   const res = await chat.invoke([message]);
   // console.log({ res });
-  expect(res.response_metadata.usage).toBeDefined();
+  expect(res.usage_metadata?.input_token_details).toBeDefined();
 });
 
 test("human message caching", async () => {
@@ -826,12 +825,12 @@ test("human message caching", async () => {
 
   const res = await model.invoke(messages);
   expect(
-    res.response_metadata.usage.cache_creation_input_tokens
+    res.usage_metadata?.input_token_details?.cache_creation
   ).toBeGreaterThan(0);
-  expect(res.response_metadata.usage.cache_read_input_tokens).toBe(0);
+  expect(res.usage_metadata?.input_token_details?.cache_read).toBe(0);
   const res2 = await model.invoke(messages);
-  expect(res2.response_metadata.usage.cache_creation_input_tokens).toBe(0);
-  expect(res2.response_metadata.usage.cache_read_input_tokens).toBeGreaterThan(
+  expect(res2.usage_metadata?.input_token_details?.cache_creation).toBe(0);
+  expect(res2.usage_metadata?.input_token_details?.cache_read).toBeGreaterThan(
     0
   );
 });

libs/langchain-anthropic/src/utils/message_outputs.ts

@@ -30,11 +30,17 @@ export function _makeMessageChunkFromAnthropicEvent(
         filteredAdditionalKwargs[key] = value;
       }
     }
-    const { input_tokens, output_tokens, ...rest } = usage ?? {};
+    // eslint-disable-next-line @typescript-eslint/no-explicit-any
+    const { input_tokens, output_tokens, ...rest }: Record<string, any> =
+      usage ?? {};
     const usageMetadata: UsageMetadata = {
       input_tokens,
       output_tokens,
       total_tokens: input_tokens + output_tokens,
+      input_token_details: {
+        cache_creation: rest.cache_creation_input_tokens,
+        cache_read: rest.cache_read_input_tokens,
+      },
     };
     return {
       chunk: new AIMessageChunk({
@@ -54,6 +60,12 @@ export function _makeMessageChunkFromAnthropicEvent(
       input_tokens: 0,
       output_tokens: data.usage.output_tokens,
       total_tokens: data.usage.output_tokens,
+      input_token_details: {
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        cache_creation: (data.usage as any).cache_creation_input_tokens,
+        // eslint-disable-next-line @typescript-eslint/no-explicit-any
+        cache_read: (data.usage as any).cache_read_input_tokens,
+      },
     };
     return {
       chunk: new AIMessageChunk({
@@ -171,6 +183,10 @@ export function anthropicResponseToChatMessages(
           input_tokens: usage.input_tokens ?? 0,
           output_tokens: usage.output_tokens ?? 0,
           total_tokens: (usage.input_tokens ?? 0) + (usage.output_tokens ?? 0),
+          input_token_details: {
+            cache_creation: usage.cache_creation_input_tokens,
+            cache_read: usage.cache_read_input_tokens,
+          },
         }
       : undefined;
   if (messages.length === 1 && messages[0].type === "text") {

yarn.lock

@@ -241,6 +241,21 @@ __metadata:
   languageName: node
   linkType: hard
 
+"@anthropic-ai/sdk@npm:^0.32.1":
+  version: 0.32.1
+  resolution: "@anthropic-ai/sdk@npm:0.32.1"
+  dependencies:
+    "@types/node": ^18.11.18
+    "@types/node-fetch": ^2.6.4
+    abort-controller: ^3.0.0
+    agentkeepalive: ^4.2.1
+    form-data-encoder: 1.7.2
+    formdata-node: ^4.3.2
+    node-fetch: ^2.6.7
+  checksum: b48982e0ce066c99afe19448c5d6b38916c2c8873fbdcd4e5116abc45bcf241359604684856bfbd20fcfe00bf544d6d6e7bcf2686a9eb198bd671839b5cd0a67
+  languageName: node
+  linkType: hard
+
 "@anthropic-ai/vertex-sdk@npm:^0.4.1":
   version: 0.4.1
   resolution: "@anthropic-ai/vertex-sdk@npm:0.4.1"
@@ -11478,7 +11493,7 @@ __metadata:
   version: 0.0.0-use.local
   resolution: "@langchain/anthropic@workspace:libs/langchain-anthropic"
   dependencies:
-    "@anthropic-ai/sdk": ^0.27.3
+    "@anthropic-ai/sdk": ^0.32.1
     "@anthropic-ai/vertex-sdk": ^0.4.1
     "@jest/globals": ^29.5.0
     "@langchain/core": "workspace:*"