Generate an Excel file from an Amazon Bedrock agent - Doconda

You’ll build an Amazon Bedrock agent that, when you ask for a spreadsheet, creates a real Excel file and gives you the link to download it. The model decides when to call Doconda; you only wire up the tools.

What you need

An AWS account with a Bedrock model enabled and its id in BEDROCK_MODEL_ID.
AWS credentials in your environment (AWS_REGION, AWS_ACCESS_KEY_ID…), like the rest of your code.
A Doconda API key in DOCONDA_API_KEY. Create it in the dashboard, under API keys.
Python 3.9+ or Node 20+.

1. Install the dependencies

pip install boto3 doconda

npm install ai @ai-sdk/amazon-bedrock @doconda/sdk

2. Connect the tools and ask for the Excel file

tools() returns Doconda’s tools (create, upload, review, edit, look up…) with their name, description and JSON Schema. Bedrock takes them as toolSpec. The agent calls create_document with format: "xlsx", waits until the file is ready and replies with the links.

import json
import os

import boto3
from doconda import Doconda

bedrock = boto3.client("bedrock-runtime")
tools = Doconda().tools()
tool_config = {"tools": [{"toolSpec": {"name": t.name, "description": t.description, "inputSchema": {"json": t.input_schema}}} for t in tools]}
messages = [{"role": "user", "content": [{"text": "Make me an Excel file with the event budget: venue, catering, sound and travel"}]}]

while True:
    res = bedrock.converse(modelId=os.environ["BEDROCK_MODEL_ID"], messages=messages, toolConfig=tool_config)
    message = res["output"]["message"]
    messages.append(message)
    if res["stopReason"] != "tool_use":
        break
    results = []
    for block in message["content"]:
        if "toolUse" in block:
            use = block["toolUse"]
            out = next(t for t in tools if t.name == use["name"]).run(use["input"])
            results.append({"toolResult": {"toolUseId": use["toolUseId"], "content": [{"text": json.dumps(out)}]}})
    messages.append({"role": "user", "content": results})

print("".join(block.get("text", "") for block in message["content"]))

import { bedrock } from "@ai-sdk/amazon-bedrock"
import { generateText, isStepCount, jsonSchema, tool } from "ai"
import { Doconda } from "@doconda/sdk"

const doconda = new Doconda()
const tools = Object.fromEntries(
  doconda.tools().map((t) => [
    t.name,
    tool({ description: t.description, inputSchema: jsonSchema(t.inputSchema), execute: (input) => t.run(input as Record<string, unknown>) }),
  ]),
)

const { text } = await generateText({
  model: bedrock(process.env.BEDROCK_MODEL_ID!),
  tools,
  stopWhen: isStepCount(5),
  prompt: "Make me an Excel file with the event budget: venue, catering, sound and travel",
})

console.log(text)

The Python loop is the Converse API one: while the model asks for a tool (stopReason == "tool_use"), you run it and send the result back. In TypeScript, the AI SDK runs that loop for you.

Result

The tool returns something like this to the model, and the model passes it on to the user in its reply:

{
  "id": "doc_01JAB3…",
  "format": "xlsx",
  "status": "ready",
  "links": { "xlsx": "https://…", "pdf": "https://…" },
  "links_note": "Links expire in 5 minutes: call get_document for fresh ones."
}

You get the Excel file (.xlsx) and a PDF version. The links expire after 5 minutes; if the user asks for them later, the agent calls get_document and gets new ones.

Next steps

Amazon Bedrock: the full guide, also with the Converse API in TypeScript.
Python SDK: every method, in case you want to call Doconda without an agent.
Complete examples on GitHub.