Skip to content

README.md

Latest commit

 

History

History
184 lines (146 loc) · 7.42 KB

README.md

File metadata and controls

184 lines (146 loc) · 7.42 KB
Raw

HiAPI Python SDK

Zero-dependency Python client for the HiAPI unified async task API (/v1/tasks) — submit an image / video / audio generation task, poll it to completion, and read the output in one call.

  • Zero runtime dependencies. Standard library only (urllib, json, hmac).
  • One-call workflow. client.tasks.run(...) submits and waits for you.
  • Typed. Dataclasses throughout, ships py.typed.
  • Webhook verification. HMAC-SHA256 signature + timestamp freshness check, built in (still deduplicate deliveries by task id).

For OpenAI-compatible chat/image endpoints, keep using the openai library with base_url="https://api.hiapi.ai/v1". This SDK focuses on what the OpenAI client can't do: the asynchronous submit → poll → retrieve lifecycle (results come back as output URLs).

Install

pip install hiapi

Requires Python 3.8+.

Quick start

from hiapi import HiAPI

client = HiAPI(api_key="sk-...")  # or set HIAPI_API_KEY

task = client.tasks.run(
    model="seedance-2-0",
    input={"prompt": "a cyan glass data center entrance", "resolution": "1080p"},
    on_update=lambda t: print("status:", t.status),
)

for out in task.output:
    print(out.type, out.url)  # e.g. "video https://cdn.hiapi.ai/tasks/..."

run() blocks until the task reaches a terminal state. It raises TaskFailed if the task fails and PollTimeout if it doesn't finish within timeout (default 600s). A poll timeout does not cancel the task — it may still finish (and bill) later; take task_id from the PollTimeout exception and retrieve() it instead of submitting the same request again.

Output URLs are temporary — they expire about 7 days after creation (each output carries expire_at). To keep an output long-term, promote it to persistent storage before it expires — see the Output Storage docs.

Lower-level control

created = client.tasks.create(
    model="seedance-2-0",
    input={"prompt": "...", "resolution": "720p"},
    callback={"url": "https://your-app.com/hiapi/callback", "when": "final"},
)
print(created.task_id)

task = client.tasks.retrieve(created.task_id)   # one status check
task = client.tasks.wait(created.task_id, poll_interval=3, timeout=900)
page = client.tasks.list(page=1, size=20)       # newest first

input fields are defined per model — see the relevant model page. Don't put callback fields inside input; pass callback separately.

Model routes

Some models expose multiple routes (e.g. ext) with different pricing or upstream capacity. Pass route instead of writing the model@route suffix:

created = client.tasks.create(
    model="gpt-image-2/text-to-image",
    route="ext",                        # preferred over model="...@ext"
    input={"prompt": "..."},
)

Omitting route (or passing "default") uses the model's default route. An unknown route fails fast with a 400 whose message lists the available routes. The legacy model="x@ext" spelling keeps working. When a task was submitted with route, its detail echoes task.route and task.model holds the resolved full name (x@ext).

Idempotent retries

Pass idempotency_key (sent as the Idempotency-Key header, ≤255 bytes) to make task submission safe to retry — retrying the same key + same body within about 24 hours returns the first task instead of creating and billing a new one (after that the key is cleaned up and the same request creates a new task):

created = client.tasks.create(
    model="seedance-2-0",
    input={"prompt": "..."},
    idempotency_key="order-8472:video",   # a stable key you derive per job
)
if created.idempotent_replay:
    print("hit the idempotency cache; no new task created")

With a key set, the SDK also retries the POST on network errors and retries 409 IDEMPOTENCY_KEY_PROCESSING (the first request is still in flight) up to the retry limit (default 2), honouring Retry-After capped at 60s. Reusing a key with a different body raises IdempotencyKeyMismatchError — that's a key-construction bug, not retryable.

Webhooks

If you set a Webhook signing key in the HiAPI console, terminal callbacks are signed. Verify them against the raw request body:

# Flask example
from flask import Flask, request
from hiapi import HiAPI, WebhookVerificationError

app = Flask(__name__)
app.config["MAX_CONTENT_LENGTH"] = 1024 * 1024  # 1 MiB — reject oversized bodies before reading
client = HiAPI(api_key="sk-...", webhook_secret="whsec_...")

@app.post("/hiapi/callback")
def callback():
    try:
        task = client.webhooks.verify(request.get_data(), request.headers)
    except WebhookVerificationError:
        return "", 400
    if task.succeeded and task.output:
        print(task.output[0].url)
    return "", 200  # ack with 2xx; HiAPI retries non-2xx

Callbacks are delivered at least once and can arrive concurrently — make your handler idempotent (e.g. upsert your own record keyed by task.task_id) and return 2xx only after processing succeeds. Duplicates are then harmless, and a failed or crashed handler is simply redelivered.

Errors

Exception When
AuthenticationError 401 — bad/missing API key
NotFoundError 404 — unknown task or not yours
InvalidRequestError INVALID_REQUEST — fix the request
ModelUnavailableError MODEL_UNAVAILABLE — retry or switch model
TaskFailedError TASK_FAILED — the submission was rejected synchronously (distinct from TaskFailed below)
TaskTimeoutError / StorageUnavailableError retryable upstream errors
ServiceUnavailableError 503 — platform busy (auto-retried)
IdempotencyKeyProcessingError 409 — same key still in flight (auto-retried; retryable)
IdempotencyKeyMismatchError 422 — key reused with a different body (not retryable)
APIError any other non-2xx response (base class — e.g. 402, 403; carries status and body)
APIConnectionError network failure (auto-retried for reads only — not a keyless create())
TaskFailed a polled task ended in status=fail
PollTimeout run()/wait() exceeded its timeout
WebhookVerificationError bad signature or stale timestamp

429/503 are retried automatically with exponential backoff (max_retries, default 2; honours Retry-After). Network errors are retried only for idempotent GETs (retrieve / list, and the polling inside wait / run). The POST that create() issues is never retried on a network failure — unless you pass idempotency_key, which makes the retry safe server-side. Without a key, an APIConnectionError from create() leaves the request in an unknown state — a task may or may not have been created (and billed). list() can help you inspect recent tasks manually, but tasks don't echo your input back, so absence from the list doesn't prove the request failed — don't retry automatically on that basis. For anything automated, submit with an idempotency_key so the retry is safe by design.

Configuration

HiAPI(
    api_key=None,                       # falls back to HIAPI_API_KEY
    base_url="https://api.hiapi.ai/v1",
    timeout=60.0,                       # per-request seconds
    max_retries=2,
    webhook_secret=None,                # falls back to HIAPI_WEBHOOK_SECRET
)

License

MIT