umpirix/sdk/python/README.md

136 lines
4.7 KiB
Markdown
Raw Permalink Normal View History

# OpenJev Python SDK
**Now part of Scribase: [scribase.com/decisions](https://scribase.com/decisions).** This
SDK itself is unchanged and still distributed as `openjev-client` — the link above is a
Jev-compatible quickstart that uses this package, not a rename or a republish.
`openjev` is an open Python client for OpenJev — a self-hostable decision engine
that speaks the same **System One** `/v1` wire protocol as the upstream TypeSafe AI API, so a
single client works unchanged against:
- a **self-hosted OpenJev** deployment (default base URL: `http://localhost:8080`)
- the hosted **TypeSafe AI API** at `https://api.typesafe.ai`
This SDK is derived from [`@typesafe-ai/typesafe-sdk-python`](https://github.com/typesafe-ai/typesafe-sdk-python) (MIT).
See [`NOTICE`](./NOTICE) for attribution. The request/response shapes (`model` + `state` +
questions-as-`dict`) and the public class names (`TypeSafeClient`, `AsyncTypeSafeClient`, …) are
kept identical so existing TypeSafe code is drop-in compatible.
## Quickstart
Install the SDK:
```
pip install openjev-client
# or: uv add openjev-client
```
> Live on PyPI as **`openjev-client`** (the importable module is `openjev`). The project
> was renamed to Umpirix; the `umpirix-client` package name is publishing soon and
> `openjev-client` will keep installing. The plain PyPI name `openjev` belongs to an
> unrelated project — do not install it. To work from source instead:
```
git clone https://github.com/caelum0x/umpirix
pip install ./umpirix/sdk/python
```
No API key is needed for a self-hosted server: the client talks to `http://localhost:8080`
by default and sends no `Authorization` header unless a key is configured. Start the server
(`npm run dev` in the repository root), then:
```python
from openjev import Choice, TypeSafeClient
with TypeSafeClient() as client:
response = client.system_one(
state={"document": "I was charged twice. Please fix this ASAP."},
questions={
"category": Choice(
instructions="What is this ticket about?",
criteria={"billing": None, "technical": None, "other": None},
),
},
)
print(response.choices["category"].choice)
```
## Configuration
Point the SDK at a self-hosted OpenJev endpoint or the upstream TypeSafe AI API by setting
environment variables or passing arguments to the client constructor.
| Purpose | Primary (OpenJev) | Back-compat alias (legacy) |
| ------------------ | -------------------------------- | --------------------------- |
| API key (optional) | `OPENJEV_API_KEY` | `TYPESAFE_API_KEY` |
| Base URL | `OPENJEV_BASE_URL` | `TYPESAFE_BASE_URL` |
| Default model | `OPENJEV_DEFAULT_MODEL` | `TYPESAFE_DEFAULT_MODEL` |
| Log level | `OPENJEV_LOG_LEVEL` | `TYPESAFE_LOG_LEVEL` |
The primary `OPENJEV_*` variable is checked first; the `TYPESAFE_*` alias is consulted only
when the primary variable is absent so that callers already using `TYPESAFE_*` continue to work
without changes.
Default base URL when no variable is set: `http://localhost:8080`.
Example — point at a self-hosted deployment (the key line is only needed if your server
issues keys):
```bash
export OPENJEV_BASE_URL="https://systemone.internal.example.com"
export OPENJEV_API_KEY="ots_live_..." # optional
```
Example — point at the upstream TypeSafe AI API using the back-compat alias:
```bash
export TYPESAFE_BASE_URL="https://api.typesafe.ai"
export TYPESAFE_API_KEY="ts_live_..."
```
```python
from openjev import TypeSafeClient
# Explicit override also works and takes precedence over the environment:
client = TypeSafeClient(
base_url="https://systemone.internal.example.com", api_key="..."
)
```
## Trust layer (OpenJev servers)
```python
from openjev import Choice, TypeSafeClient
with TypeSafeClient() as client:
questions = {
"intent": Choice(
instructions="What does the customer want?",
criteria={"refund": None, "status": None},
abstain=True,
)
}
result = client.system_one(
"Where is my order?",
questions,
model="auto", # local first, LLM only for low-confidence questions
calibration_id="cal_...", # confidence fitted on your own labels
)
intent = result.choices["intent"]
if intent.abstained:
print("out of scope", intent.ood)
# Shadow mode: compare with your production LLM or human labels.
client.shadow.run(
"Where is my order?",
questions,
incumbent={"kind": "label", "answers": {"intent": "status"}},
)
summary = client.shadow.summary(days=7)
```
## License
MIT — see [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE).