velmhq 0.3.0

Python SDK for the Velm API.

velmhq

Python SDK for the Velm API.

Status: pre-release. The public surface is stable in shape but the underlying API is still evolving. Pin a version until the 1.0.0 release.

Install

pip install velmhq
# or
uv add velmhq

The PyPI distribution name is velmhq because the bare velm is already published by an unrelated project.

Usage

from velmhq import Velm

velm = Velm(api_key="vk_test_...")

# Invoke an agent and wait for the result.
run = velm.run("support-bot", {"question": "How do I reset my password?"})
print(run.text)

# Capture a thumbs-down to feed the improvement loop.
velm.feedback.capture(
    run_id=run.id,
    feedback_key=run.id,
    event_type="thumbs_down",
    conversation_text=transcript,
    original_response=bad_response,
    metadata={"surface": "sdk"},
)

The async client mirrors the sync surface:

import asyncio
from velmhq import AsyncVelm

async def main() -> None:
    async with AsyncVelm(api_key="vk_test_...") as velm:
        run = await velm.run("support-bot", {"question": "..."})
        print(run.text)

asyncio.run(main())

Multi-turn conversations

Pass the same conversation_id across runs and the agent replays the prior turns, so it has the full context. The id is caller-minted - any stable string. Omit it for a one-shot run.

import uuid

conversation_id = str(uuid.uuid4())

turn1 = velm.runs.create("support-bot",
    input={"message": "I was double-charged last month."},
    conversation_id=conversation_id)

turn2 = velm.runs.create("support-bot",
    input={"message": "and what about this month?"},
    conversation_id=conversation_id)

# Inspect every run in the conversation.
for run in velm.runs.list(conversation_id=conversation_id):
    print(run.id, run.status, run.created_at)

Errors

The SDK raises typed exceptions on non-2xx responses. Branch on the class, not the message string.

from velmhq import (
    AuthenticationError,
    BillingError,
    RateLimitError,
    VelmError,
)

try:
    velm.feedback.capture(...)
except RateLimitError as err:
    print(f"rate limited; retry after {err.retry_after_ms}ms")
except BillingError as err:
    print(err.message)  # points at the billing settings page
except AuthenticationError:
    print("invalid API key")
except VelmError as err:
    print(f"{err.code}: {err.message}")

Testing

velmhq.testing.MockVelm is a scriptable fake for unit tests in your own code. It runs the real SDK code path under a fake transport, so retries, idempotency keys, and error mapping all behave as in production.

from velmhq.testing import MockVelm

def test_my_handler() -> None:
    velm = MockVelm()
    velm.respond_next({"feedbackId": "fb_1"})

    my_handler(velm)

    assert len(velm.calls) == 1
    assert velm.calls[0].url.endswith("/feedback")

Runtime support

Works on CPython 3.10 through 3.13. Uses httpx for HTTP and pydantic v2 for typed models. Both sync (Velm) and async (AsyncVelm) clients are provided; pick whichever fits your application's runtime.

Documentation

Full reference at velm.run/docs.

License

MIT. See LICENSE.