Merge pull request #37 from langchain-ai/brace/custom-parsing-anthropic

fix: Custom tool call parsing for anthropic

agent/agent.ts

@@ -7,10 +7,12 @@ import { ChatOpenAI } from "@langchain/openai";
 import { tripPlannerGraph } from "./trip-planner";
 import { formatMessages } from "./utils/format-messages";
 import { graph as openCodeGraph } from "./open-code";
+import { graph as orderPizzaGraph } from "./pizza-orderer";
 
 const allToolDescriptions = `- stockbroker: can fetch the price of a ticker, purchase/sell a ticker, or get the user's portfolio
 - tripPlanner: helps the user plan their trip. it can suggest restaurants, and places to stay in any given location.
-- openCode: can write code for the user. call this tool when the user asks you to write code`;
+- openCode: can write code for the user. call this tool when the user asks you to write code
+- orderPizza: can order a pizza for the user`;
 
 async function router(
   state: GenerativeUIState,
@@ -21,7 +23,7 @@ ${allToolDescriptions}
 `;
   const routerSchema = z.object({
     route: z
-      .enum(["stockbroker", "tripPlanner", "openCode", "generalInput"])
+      .enum(["stockbroker", "tripPlanner", "openCode", "orderPizza", "generalInput"])
       .describe(routerDescription),
   });
   const routerTool = {
@@ -75,7 +77,7 @@ Please pick the proper route based on the most recent message, in the context of
 
 function handleRoute(
   state: GenerativeUIState,
-): "stockbroker" | "tripPlanner" | "openCode" | "generalInput" {
+): "stockbroker" | "tripPlanner" | "openCode" | "orderPizza" | "generalInput" {
   return state.next;
 }
 
@@ -107,18 +109,21 @@ const builder = new StateGraph(GenerativeUIAnnotation)
   .addNode("stockbroker", stockbrokerGraph)
   .addNode("tripPlanner", tripPlannerGraph)
   .addNode("openCode", openCodeGraph)
+  .addNode("orderPizza", orderPizzaGraph)
   .addNode("generalInput", handleGeneralInput)
 
   .addConditionalEdges("router", handleRoute, [
     "stockbroker",
     "tripPlanner",
     "openCode",
+    "orderPizza",
     "generalInput",
   ])
   .addEdge(START, "router")
   .addEdge("stockbroker", END)
   .addEdge("tripPlanner", END)
   .addEdge("openCode", END)
+  .addEdge("orderPizza", END)
   .addEdge("generalInput", END);
 
 export const graph = builder.compile();

agent/pizza-orderer/index.ts

@@ -0,0 +1,85 @@
+import { ChatAnthropic } from "@langchain/anthropic";
+import { Annotation, END, START, StateGraph } from "@langchain/langgraph";
+import { GenerativeUIAnnotation } from "../types";
+import { z } from "zod";
+import { AIMessage, ToolMessage } from "@langchain/langgraph-sdk";
+import { v4 as uuidv4 } from "uuid";
+
+const PizzaOrdererAnnotation = Annotation.Root({
+  messages: GenerativeUIAnnotation.spec.messages,
+})
+
+async function sleep(ms = 5000) {
+  return new Promise((resolve) => setTimeout(resolve, ms));
+}
+
+const workflow = new StateGraph(PizzaOrdererAnnotation)
+  .addNode("findStore", async (state) => {
+    const findShopSchema = z.object({
+      location: z.string().describe("The location the user is in. E.g. 'San Francisco' or 'New York'"),
+      pizza_company: z.string().optional().describe("The name of the pizza company. E.g. 'Dominos' or 'Papa John's'. Optional, if not defined it will search for all pizza shops"),
+    }).describe("The schema for finding a pizza shop for the user")
+    const model = new ChatAnthropic({ model: "claude-3-5-sonnet-latest", temperature: 0 }).withStructuredOutput(findShopSchema, {
+      name: "find_pizza_shop",
+      includeRaw: true,
+    })
+
+    const response = await model.invoke([
+      {
+        role: "system",
+        content: "You are a helpful AI assistant, tasked with extracting information from the conversation between you, and the user, in order to find a pizza shop for them."
+      },
+      ...state.messages,
+    ])
+
+    await sleep();
+
+    const toolResponse: ToolMessage = {
+      type: "tool",
+      id: uuidv4(),
+      content: "I've found a pizza shop at 1119 19th St, San Francisco, CA 94107. The phone number for the shop is 415-555-1234.",
+      tool_call_id: (response.raw as unknown as AIMessage).tool_calls?.[0].id ?? "",
+    }
+    
+    return {
+      messages: [response.raw, toolResponse]
+    }
+  })
+  .addNode("orderPizza", async (state) => {
+    await sleep(1500);
+
+    const placeOrderSchema = z.object({
+      address: z.string().describe("The address of the store to order the pizza from"),
+      phone_number: z.string().describe("The phone number of the store to order the pizza from"),
+      order: z.string().describe("The full pizza order for the user"),
+    }).describe("The schema for ordering a pizza for the user")
+    const model = new ChatAnthropic({ model: "claude-3-5-sonnet-latest", temperature: 0 }).withStructuredOutput(placeOrderSchema, {
+      name: "place_pizza_order",
+      includeRaw: true,
+    })
+
+    const response = await model.invoke([
+      {
+        role: "system",
+        content: "You are a helpful AI assistant, tasked with placing an order for a pizza for the user."
+      },
+      ...state.messages,
+    ])
+
+    const toolResponse: ToolMessage = {
+      type: "tool",
+      id: uuidv4(),
+      content: "Pizza order successfully placed.",
+      tool_call_id: (response.raw as unknown as AIMessage).tool_calls?.[0].id ?? "",
+    }
+    
+    return {
+      messages: [response.raw, toolResponse]
+    }
+  })
+  .addEdge(START, "findStore")
+  .addEdge("findStore", "orderPizza")
+  .addEdge("orderPizza", END)
+
+export const graph = workflow.compile()
+graph.name = "Order Pizza Graph";

agent/types.ts

@@ -13,7 +13,7 @@ export const GenerativeUIAnnotation = Annotation.Root({
   >({ default: () => [], reducer: uiMessageReducer }),
   timestamp: Annotation<number>,
   next: Annotation<
-    "stockbroker" | "tripPlanner" | "openCode" | "generalInput"
+    "stockbroker" | "tripPlanner" | "openCode" | "orderPizza" | "generalInput"
   >(),
 });
 

package.json

@@ -16,6 +16,7 @@
     "@assistant-ui/react-markdown": "^0.8.0",
     "@assistant-ui/react-syntax-highlighter": "^0.7.2",
     "@faker-js/faker": "^9.5.1",
+    "@langchain/anthropic": "^0.3.15",
     "@langchain/core": "^0.3.41",
     "@langchain/google-genai": "^0.1.10",
     "@langchain/langgraph": "^0.2.49",

src/components/thread/messages/ai.tsx

@@ -1,5 +1,6 @@
+import { parsePartialJson } from "@langchain/core/output_parsers";
 import { useStreamContext } from "@/providers/Stream";
-import { Checkpoint, Message } from "@langchain/langgraph-sdk";
+import { AIMessage, Checkpoint, Message } from "@langchain/langgraph-sdk";
 import { getContentString } from "../utils";
 import { BranchSwitcher, CommandBar } from "./shared";
 import { MarkdownText } from "../markdown-text";
@@ -7,6 +8,7 @@ import { LoadExternalComponent } from "@langchain/langgraph-sdk/react-ui";
 import { cn } from "@/lib/utils";
 import { ToolCalls, ToolResult } from "./tool-calls";
 import { StringParam, useQueryParam } from "use-query-params";
+import { MessageContentComplex } from "@langchain/core/messages";
 
 function CustomComponent({
   message,
@@ -44,6 +46,28 @@ function CustomComponent({
   );
 }
 
+function parseAnthropicStreamedToolCalls(content: MessageContentComplex[]): AIMessage["tool_calls"] {
+  const toolCallContents = content.filter((c) => c.type === "tool_use" && c.id);
+
+  return toolCallContents.map((tc) => {
+    const toolCall = tc as Record<string, any>
+    let json: Record<string, any> = {};
+    if (toolCall?.input) {
+      try {
+        json = parsePartialJson(toolCall.input) ?? {}
+      } catch {
+        // Pass
+      }
+    }
+    return {
+      name: toolCall.name ?? "",
+      id: toolCall.id ?? "",
+      args: json,
+      type: "tool_call",
+    }
+  })
+}
+
 export function AssistantMessage({
   message,
   isLoading,
@@ -58,11 +82,14 @@ export function AssistantMessage({
   const thread = useStreamContext();
   const meta = thread.getMessagesMetadata(message);
   const parentCheckpoint = meta?.firstSeenState?.parent_checkpoint;
+  const anthropicStreamedToolCalls = Array.isArray(message.content) ? parseAnthropicStreamedToolCalls(message.content) : undefined;
 
   const hasToolCalls =
-    "tool_calls" in message &&
+    ("tool_calls" in message &&
       message.tool_calls &&
-    message.tool_calls.length > 0;
+      message.tool_calls.length > 0);
+  const toolCallsHaveContents = hasToolCalls && message.tool_calls?.some((tc) => tc.args && Object.keys(tc.args).length > 0);
+  const hasAnthropicToolCalls = !!anthropicStreamedToolCalls?.length;
   const isToolResult = message.type === "tool";
 
   return (
@@ -76,7 +103,9 @@ export function AssistantMessage({
               <MarkdownText>{contentString}</MarkdownText>
             </div>
           )}
-          {hasToolCalls && <ToolCalls toolCalls={message.tool_calls} />}
+          {(hasToolCalls && toolCallsHaveContents && <ToolCalls toolCalls={message.tool_calls} />) ||
+           (hasAnthropicToolCalls && <ToolCalls toolCalls={anthropicStreamedToolCalls} />) ||
+           (hasToolCalls && <ToolCalls toolCalls={message.tool_calls} />)}
           <CustomComponent message={message} thread={thread} />
           <div
             className={cn(