Accept document inputs

Resolve URL, base64, or volume inputs with bounded downloads, checked targets, and path containment.

resolve_input_bytes() reads exactly one input source and returns the document bytes with a source label. A worker can then use detect_format() for a coarse container check before handing the bytes to its engine.

Resolve an inline document

import asyncio

from runpod_doc_worker.transport.io import (
    detect_format,
    resolve_input_bytes,
)


async def main() -> None:
    file_bytes, source = await resolve_input_bytes(
        {"file_b64": "JVBERi0xLjcK"}
    )
    print(source)                    # b64
    print(detect_format(file_bytes)) # pdf


asyncio.run(main())

The input mapping must contain exactly one truthy key:

file_url
file_b64
volume_path

Input limits

Source	Limit	Notes
`file_b64`	20 MiB decoded	Encoded length is checked before allocating the decoded copy
`file_url`	200 MiB	`Content-Length` and streamed bytes are both checked
`volume_path`	No package size cap	The file is read fully into memory

URL fetches have a 120-second inactivity timeout and a 300-second wall-clock budget. HTTP and HTTPS are the only accepted URL schemes.

URL target checks

Direct connections resolve the hostname and reject non-public addresses. The checked address is also used when the socket opens, which closes the gap between validating a hostname and connecting to it. Redirect requests go through the same check.

Set <PREFIX>_ALLOW_LOCAL_FETCH=1 only for controlled development or private-network cases. It removes the public-routability restriction and is unsafe when an untrusted caller chooses the URL.

When an environment proxy handles a hostname, the proxy resolves the final destination and must enforce its own destination policy. Literal IP targets still receive package-level checks.

Do not log the raw URL label

The returned source label for a URL contains the original URL. Use telemetry_source_kind(source) to record url, b64, volume, or unknown without copying the URL into telemetry.

Volume containment

volume_path must be absolute, resolve to a file, and remain under an allowed root after resolving .. components and symlinks. Use the narrowest roots that fit the worker:

ACME_VOLUME_ROOTS=/runpod-volume/incoming,/worker/fixtures

Format detection

detect_format() returns pdf, image, ooxml, or unknown from magic bytes.

image covers PNG, JPEG, GIF, BMP, TIFF, and WebP.
ooxml means a ZIP container. It does not prove that the file is DOCX, PPTX, or XLSX.

The engine remains responsible for accepting or rejecting the detected content.

Review the security model before enabling URL or volume inputs on a public endpoint.