Model I/O with files

Send file content in the format your model accepts: JSON values, base64-encoded bytes, or a URL that your model downloads. Your Truss defines the payload fields and output format. You need an active deployment, its model ID in BASETEN_MODEL_ID, and a Baseten API key in BASETEN_API_KEY. The JSON and base64 examples use an identity Truss that returns its input unchanged. Install requests with pip install requests for the Python clients.

Files as input

Send a file with JSON-serializable content

Save {"prompt": "Hello, world!"} as input.json, then choose a client.

Baseten CLI
Python

To send JSON from a file:Run the following command:

baseten model predict --model-id YOUR_MODEL_ID --file input.json

{"prompt": "Hello, world!"}

To send JSON from a file:Run the following script:

import os
import json
import requests

with open("input.json") as source:
    payload = json.load(source)
response = requests.post(
    f"https://model-{os.environ['BASETEN_MODEL_ID']}.api.baseten.co/environments/production/predict",
    headers={"Authorization": f"Bearer {os.environ['BASETEN_API_KEY']}"},
    json=payload,
    timeout=60,
)
response.raise_for_status()
print(response.json())

{'prompt': 'Hello, world!'}

Send a file with non-serializable content

The CLI’s --file option reads JSON. For audio, images, or other bytes, base64-encode the file if your model expects a base64 field. Use Binary I/O when your Truss accepts MessagePack instead. To round-trip file bytes through JSON: Run the following script against the identity Truss. It creates a UTF-8 file, sends it in a base64 field, and saves the returned bytes. Replace file with the field expected by your model:

import base64
import os
from pathlib import Path
import requests

source = Path("input.txt")
source.write_text("Café from Baseten.", encoding="utf-8")
response = requests.post(
    f"https://model-{os.environ['BASETEN_MODEL_ID']}.api.baseten.co/environments/production/predict",
    headers={"Authorization": f"Bearer {os.environ['BASETEN_API_KEY']}"},
    json={"file": base64.b64encode(source.read_bytes()).decode("ascii")},
    timeout=60,
)
response.raise_for_status()
output = Path("output.txt")
output.write_bytes(base64.b64decode(response.json()["file"]))
print("Bytes match:", output.read_bytes() == source.read_bytes())
print(output.read_text(encoding="utf-8"))

Bytes match: True
Café from Baseten.

Download files inside your model

To implement an input such as {"url": "https://example.com/recording.wav"}, download the file in preprocess(). The following methods belong in an existing Whisper Truss that already loads self._model; they illustrate the model implementation rather than a complete deployment. Add requests to the model’s requirements:

model/model.py

from tempfile import NamedTemporaryFile
import requests

# Add these methods to your existing Model class.
def preprocess(self, model_input):
    response = requests.get(model_input["url"], timeout=30)
    response.raise_for_status()
    return {"content": response.content}


def predict(self, model_input):
    with NamedTemporaryFile() as audio:
        audio.write(model_input["content"])
        audio.flush()
        result = self._model.transcribe(audio.name)
    return {"text": result["text"]}

Flushing the file makes the written bytes visible before Whisper opens its path. The temporary file is removed after transcription. Deploy the changed Truss before sending URL inputs. For a complete transcription workflow that uploads a local file, use the speech quickstart.

Files as output

Save model output to a local file

Baseten CLI
Python

To save the JSON response:Run the prediction and inspect the saved file:

baseten model predict --model-id YOUR_MODEL_ID --data '{"prompt":"Hello, world!"}' > output.json
cat output.json

{"prompt": "Hello, world!"}

To save the JSON response:Continue call_file.py after checking the response status:

with open("output.json", "w") as output:
    json.dump(response.json(), output)
with open("output.json") as output:
    print(json.load(output))

{'prompt': 'Hello, world!'}

Decode base64 output before writing image or audio bytes, as in the round-trip example above. For more information, see the image generation example.

Get started

Model APIs

Baseten Grid

Inference

Frontier Gateway

Training

Sandboxes

Development

Deployment

Engines

Model Labs

Organization

Hosting options

Observability

Troubleshooting

Model I/O with files

Files as input

Send a file with JSON-serializable content

Send a file with non-serializable content

Download files inside your model

Files as output

Save model output to a local file

​Files as input

​Send a file with JSON-serializable content

​Send a file with non-serializable content

​Download files inside your model

​Files as output

​Save model output to a local file

Files as input

Send a file with JSON-serializable content

Send a file with non-serializable content

Download files inside your model

Files as output

Save model output to a local file