feat(langchain): add a big tool middleware

Author: christian-bromann

examples/src/createAgent/middleware/bigTool.ts

@@ -0,0 +1,100 @@
+import * as z from "zod";
+import { createAgent, HumanMessage, AIMessage, tool } from "langchain";
+import { bigToolMiddleware } from "langchain/middleware";
+import { ChatAnthropic } from "@langchain/anthropic";
+
+/**
+ * Capture the raw requests to verify the tools are being selected correctly
+ */
+const requestBodies: any[] = [];
+const llm = new ChatAnthropic({
+  apiKey: process.env.ANTHROPIC_API_KEY,
+  model: "claude-sonnet-4-20250514",
+  clientOptions: {
+    fetch: (url, options) => {
+      requestBodies.push(JSON.parse(options?.body as string));
+      return fetch(url, options);
+    },
+  },
+});
+
+// Example usage
+const agent = createAgent({
+  llm,
+  tools: [
+    tool(
+      ({ a, b, operation }: { a: number; b: number; operation: string }) => {
+        if (operation === "add") {
+          return a + b;
+        } else if (operation === "multiply") {
+          return a * b;
+        } else if (operation === "subtract") {
+          return a - b;
+        } else if (operation === "divide") {
+          return a / b;
+        } else {
+          return "Unknown operation";
+        }
+      },
+      {
+        name: "calculator",
+        description: "Perform basic math operations",
+        schema: z.object({
+          a: z.number(),
+          b: z.number(),
+          operation: z.enum(["add", "multiply", "subtract", "divide"]),
+        }),
+      }
+    ),
+    tool(
+      () => {
+        return "Database query result: 42";
+      },
+      {
+        name: "database",
+        description: "Query a database",
+        schema: z.object({
+          query: z.string(),
+        }),
+      }
+    ),
+  ],
+  middleware: [
+    bigToolMiddleware({
+      strategy: "custom",
+      customSelector: (tools, query) => {
+        if (query.includes("calculator")) {
+          return tools.filter((tool) => tool.name === "calculator");
+        } else if (query.includes("database")) {
+          return tools.filter((tool) => tool.name === "database");
+        }
+
+        return [];
+      },
+      maxTools: 1,
+    }),
+  ] as const,
+});
+
+// Usage example with a long chat history for testing caching
+const result = await agent.invoke({
+  messages: [new HumanMessage("Calculate 10 + 20")],
+});
+console.log(
+  "Calculate 10 + 20\nAgent response:",
+  result.messages.at(-1)?.content,
+  `\nAll available tools: ${requestBodies
+    .pop()
+    .tools.map((tool: { name: string }) => tool.name)}`
+);
+
+const result2 = await agent.invoke({
+  messages: [new HumanMessage("Query the database for 42")],
+});
+console.log(
+  "\nQuery the database for 42\nAgent response:",
+  result2.messages.at(-1)?.content,
+  `\nAll available tools: ${requestBodies
+    .pop()
+    .tools.map((tool: { name: string }) => tool.name)}`
+);