fix: calling tools with large inputs

.changeset/nasty-peas-exercise.md

@@ -0,0 +1,5 @@
+---
+"llamaindex": patch
+---
+
+Show error message if agent tool is called with partial JSON

examples/agent/large_toolcall.ts

@@ -0,0 +1,64 @@
+import { FunctionTool, OpenAI, OpenAIAgent } from "llamaindex";
+
+const csvData =
+  "TITLE,RELEASE_YEAR,SCORE,NUMBER_OF_VOTES,DURATION,MAIN_GENRE,MAIN_PRODUCTION\nDavid Attenborough: A Life on Our Planet,2020,9,31180,83,documentary,GB\nInception,2010,8.8,2268288,148,scifi,GB\nForrest Gump,1994,8.8,1994599,142,drama,US\nAnbe Sivam,2003,8.7,20595,160,comedy,IN\nBo Burnham: Inside,2021,8.7,44074,87,comedy,US\nSaving Private Ryan,1998,8.6,1346020,169,drama,US\nDjango Unchained,2012,8.4,1472668,165,western,US\nDangal,2016,8.4,180247,161,action,IN\nBo Burnham: Make Happy,2016,8.4,14356,60,comedy,US\nLouis C.K.: Hilarious,2010,8.4,11973,84,comedy,US\nDave Chappelle: Sticks & Stones,2019,8.4,25687,65,comedy,US\n3 Idiots,2009,8.4,385782,170,comedy,IN\nBlack Friday,2004,8.4,20611,143,crime,IN\nSuper Deluxe,2019,8.4,13680,176,thriller,IN\nWinter on Fire: Ukraine's Fight for Freedom,2015,8.3,17710,98,documentary,UA\nOnce Upon a Time in America,1984,8.3,342335,229,drama,US\nTaxi Driver,1976,8.3,795222,113,crime,US\nLike Stars on Earth,2007,8.3,188234,165,drama,IN\nBo Burnham: What.,2013,8.3,11488,60,comedy,US\nFull Metal Jacket,1987,8.3,723306,116,drama,GB\nWarrior,2011,8.2,463276,140,drama,US\nDrishyam,2015,8.2,79075,163,thriller,IN\nQueen,2014,8.2,64805,146,drama,IN\nPaan Singh Tomar,2012,8.2,35888,135,drama,IN";
+
+const userQuestion = "which are the best comedies after 2010?";
+
+(async () => {
+  // The agent will succeed if we increase `maxTokens` to 1024
+  const llm = new OpenAI({ model: "gpt-4-turbo", maxTokens: 256 });
+
+  type Input = {
+    code: string;
+  };
+  // initiate fake code interpreter
+  const interpreterTool = FunctionTool.from<Input>(
+    ({ code }) => {
+      console.log(
+        `To answer the user's question, call the following code:\n${code}`,
+      );
+      return code;
+    },
+    {
+      name: "interpreter",
+      description:
+        "Execute python code in a Jupyter notebook cell and return any result, stdout, stderr, display_data, and error.",
+      parameters: {
+        type: "object",
+        properties: {
+          code: {
+            type: "string",
+            description: "The python code to execute in a single cell.",
+          },
+        },
+        required: ["code"],
+      },
+    },
+  );
+
+  const systemPrompt =
+    "You are a Python interpreter.\n        - You are given tasks to complete and you run python code to solve them.\n        - The python code runs in a Jupyter notebook. Every time you call $(interpreter) tool, the python code is executed in a separate cell. It's okay to make multiple calls to $(interpreter).\n        - Display visualizations using matplotlib or any other visualization library directly in the notebook. Shouldn't save the visualizations to a file, just return the base64 encoded data.\n        - You can install any pip package (if it exists) if you need to but the usual packages for data analysis are already preinstalled.\n        - You can run any python code you want in a secure environment.";
+
+  const agent = new OpenAIAgent({
+    llm,
+    tools: [interpreterTool],
+    systemPrompt,
+    verbose: true,
+  });
+
+  console.log(`User question: ${userQuestion}\n`);
+
+  await agent.chat({
+    message: [
+      {
+        type: "text",
+        text: userQuestion,
+      },
+      {
+        type: "text",
+        text: `Use data from following CSV raw contents:\n${csvData}`,
+      },
+    ],
+  });
+})();

packages/core/src/agent/utils.ts

@@ -16,10 +16,25 @@ export async function callTool(
   toolCall: ToolCall | PartialToolCall,
   logger: Logger,
 ): Promise<ToolOutput> {
-  const input: JSONObject =
-    typeof toolCall.input === "string"
-      ? JSON.parse(toolCall.input)
-      : toolCall.input;
+  let input: JSONObject;
+  if (typeof toolCall.input === "string") {
+    try {
+      input = JSON.parse(toolCall.input);
+    } catch (e) {
+      const output = `Tool ${toolCall.name} can't be called. Input is not a valid JSON object.`;
+      logger.error(
+        `${output} Try increasing the maxTokens parameter of your LLM. Invalid Input: ${toolCall.input}`,
+      );
+      return {
+        tool,
+        input: {},
+        output,
+        isError: true,
+      };
+    }
+  } else {
+    input = toolCall.input;
+  }
   if (!tool) {
     logger.error(`Tool ${toolCall.name} does not exist.`);
     const output = `Tool ${toolCall.name} does not exist.`;