Perplexity Decisions API as a publish gate for Sume output

You can put a decision model after a Sume Format run: when the run finishes, your receiver downloads the image from primary_output_url, sends it to Perplexity's Decisions API with one yes/no question, and publishes only if the returned probability clears a threshold you choose. Sume does not call Perplexity for you; this is a step in your own code, after the webhook.

The check is independent of the run, which is the point. A Format's output_schema can make the generating agent report a boolean about its own work, but an agent grading itself is weaker evidence than a separate model looking at the file.

What does the Decisions API take and return?

Perplexity's quickstart describes a single endpoint, POST https://api.perplexity.ai/v1/decisions, with Bearer auth and the model pplx-decider-v1-27b. You pass a state and a set of named questions; a noul question answers with a probability of yes between 0 and 1.

What does the gate look like in code?

This takes the receipt's primary_output_url, which for an image Format is a durable media.sume.com URL, and returns whether to publish. It refuses to guess when the file type is not one the API lists.

import base64, os, requests

OK_TYPES = {"image/png", "image/jpeg", "image/webp"}

def gate(image_url, question, threshold=0.9):
    img = requests.get(image_url, timeout=60)
    img.raise_for_status()
    kind = img.headers.get("content-type", "").split(";")[0]
    if kind not in OK_TYPES:
        return "review"
    data_url = f"data:{kind};base64," + base64.b64encode(img.content).decode()
    body = {
        "model": "pplx-decider-v1-27b",
        "state": [{"type": "image_url", "image_url": {"url": data_url}}],
        "questions": {"ok": {"type": "noul", "instructions": question, "criteria": "Answer yes only if it is clearly true."}},
    }
    r = requests.post(
        "https://api.perplexity.ai/v1/decisions",
        json=body,
        headers={"Authorization": "Bearer " + os.environ["PERPLEXITY_API_KEY"]},
        timeout=60,
    )
    if r.status_code != 200:
        return "review"
    p = r.json()["answers"]["ok"]["noul"]
    return "publish" if p >= threshold else "review"

Where does it sit in the webhook flow?

The gate belongs in the worker that handles the webhook, after the delivery has been acknowledged.

• Verify the signature and answer 2xx first; Sume gives a receiver 10 seconds, so queue the gate instead of running it inline.
• Fetch the receipt if payload is null, which happens when the delivery would exceed 1 MiB; error.result_url points to it.
• Branch on outcome. Only ok has a usable output; on degraded there is media but no structured output, so send it straight to review.
• Run gate() on primary_output_url. A publish goes out; anything else goes to a person with the receipt id.

What does this not cover?

The quickstart lists images, not video, so a clip needs a still frame that you extract yourself, for example with ffmpeg, before the call. A single frame cannot vouch for the other 29 seconds. Treat the check as a cheap filter for obvious misses such as a missing logo or the wrong product, not as approval of the whole video.

Every failure path in the function returns review, never publish. A 429 or a 504 from the decision API, a wrong file type and a network error all leave the asset waiting for a person, which is the safe default for an automated pipeline. Pick the threshold by running the gate over assets you have already judged by eye, and keep each question about one visible fact, such as whether the logo appears.