Metadata-Version: 2.4
Name: ehs-llm-client
Version: 0.2.1
Summary: Async client for OpenAI, Azure OpenAI, Microsoft Foundry, Google Gemini and TypeSafe decision models
Author-email: Andersen Huang <andersen.huang@ehsanalytics.com>
License-Expression: MIT
Requires-Python: >=3.10
Description-Content-Type: text/markdown
Requires-Dist: openai>=1.97.0
Requires-Dist: httpx[http2]
Requires-Dist: google-genai
Requires-Dist: python-dotenv
Requires-Dist: python-dateutil

# llm-client

[![PyPI version](https://img.shields.io/pypi/v/ehs-llm-client)](https://pypi.org/project/ehs-llm-client/)
[![Python Version](https://img.shields.io/pypi/pyversions/ehs-llm-client)](https://www.python.org/downloads/)

`ehs-llm-client` is a unified **async Python client** for **OpenAI**, **Azure OpenAI**,
**Microsoft Foundry (AIF)**, **Google Gemini**, and **TypeSafe/Jev**.
It supports generative responses and native typed decisions. Provider batch APIs
remain available for supported generative deployments.

---

## Features

- Async support for multiple LLM providers  
- Unified interface across OpenAI, Azure, and Google Gemini  
- Structured JSON responses via schemas  
- Native `choice`, `noul` and `score` decisions via Foundry and TypeSafe
- Retry and timeout handling  
- Batch processing support  
- Environment variable configuration for API keys  
- Easy integration into existing projects or monorepos  

---

## Installation

```bash
# Install from PyPI
pip install ehs-llm-client
```

Or install editable version during development:

```bash
# From project root
pip install -e .
```

---

## Configuration

`ehs-llm-client` supports **three configuration modes**:

### 1️⃣ Config file (`.cfg`)

Example: `llmconfig.cfg`

```ini
[default_settings]
provider = openai
model = gpt-4.1-mini

[prod]
provider = openai
model = gpt-4.1
model_batch = gpt-4.1-mini
```

Usage:

```python
from ehs_llm_client import LLM

llm = LLM("prod", config_file_path="llmconfig.cfg")
```

---

### 2️⃣ Config dictionary (for tests / CI)

```python
from ehs_llm_client import LLM

llm = LLM(
    "default",
    config={
        "default": {
            "provider": "openai",
            "model": "gpt-4.1-mini"
        }
    }
)
```

---

### 3️⃣ Environment variable fallback

```bash
export LLM_PROVIDER=openai
export LLM_MODEL=gpt-4.1-mini
export OPENAI_API_KEY=sk-...
```

```python
llm = LLM("default")
```

### Azure OpenAI with GPT-6

Deploy your chosen GPT-6 model in Azure, then use its **deployment name** as `model`:

```ini
[azure]
provider = azure_openai
model = my-gpt6-deployment
base_url = https://YOUR-RESOURCE.openai.azure.com/
```

Set `AZURE_OPENAI_API_KEY` and initialize the client:

```python
llm = LLM("azure", config_file_path="llmconfig.cfg")
```

You can also configure everything with environment variables:

```powershell
$env:LLM_PROVIDER = "azure_openai"
$env:LLM_MODEL = "my-gpt6-deployment"
$env:AZURE_OPENAI_ENDPOINT = "https://YOUR-RESOURCE.openai.azure.com/"
$env:AZURE_OPENAI_API_KEY = "YOUR-AZURE-API-KEY"
```

Then use `LLM("default")`. `LLM_BASE_URL` or a configured `base_url` takes precedence
over `AZURE_OPENAI_ENDPOINT`. Both a resource URL and a full `/openai/v1/` URL are
accepted, including endpoints on `services.ai.azure.com`. The client uses Azure's
v1 API; existing `api_version` settings are retained for compatibility and ignored.
See [Azure Responses API documentation](https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/responses)
for supported models and regions.

### Microsoft Foundry (AIF) and TypeSafe

Version 0.2.0 adds two providers. Existing providers, credential precedence and
the `get_response_async()` positional arguments and return tuple are preserved.

| Provider | `api` | Method | Endpoint |
| --- | --- | --- | --- |
| `azure_foundry` | `responses` (default) | `get_response_async()` | `/openai/v1/responses` |
| `azure_foundry` | `systemone` | `decide_async()` | `/providers/microsoft/v1/systemone` |
| `typesafe` | `systemone` (default) | `decide_async()` | `https://api.typesafe.ai/v1/systemone` |

Select the API explicitly for Foundry decision deployments; deployment names do
not determine the protocol. This release supports these two APIs, rather than
every model/API in the Foundry catalog.

```ini
[foundry_gpt]
provider = azure_foundry
api = responses
model = gpt-6-luna
base_url = https://YOUR-RESOURCE.services.ai.azure.com
default_reasoning_effort = low

[foundry_decision]
provider = azure_foundry
api = systemone
model = decision-1
base_url = https://YOUR-RESOURCE.services.ai.azure.com

[typesafe_jev]
provider = typesafe
model = jev-latest
```

For Foundry, `model` is the **deployment name**. For TypeSafe, it is the model
name or alias. Foundry uses its own `AZURE_FOUNDRY_API_KEY` and
`AZURE_FOUNDRY_ENDPOINT`; it never falls back to Azure OpenAI credentials.
A configured `base_url` takes precedence over `LLM_BASE_URL`, then the
provider-specific endpoint variable. TypeSafe defaults to `https://api.typesafe.ai`;
an explicit `base_url` or environment-only `LLM_BASE_URL` can override it.

Foundry expects a resource endpoint, not an `/api/projects/...` URL. Responses
also accepts a full `/openai/v1/` endpoint. System One accepts the resource root,
its corresponding versioned API prefix, or its full System One endpoint.
These endpoints must use HTTPS and contain no query string, fragment or embedded
credentials. Selecting `azure_foundry` does not itself establish project billing
attribution; that depends on the Azure endpoint and service capabilities.

For environment-only configuration, set `LLM_PROVIDER`, `LLM_MODEL`, and
`LLM_API=systemone` for Foundry decision models, together with its endpoint/key.
TypeSafe needs `LLM_PROVIDER=typesafe`, `LLM_MODEL=jev-latest`, and
`TYPESAFE_API_KEY`.

See the [Microsoft Decision API](https://learn.microsoft.com/en-us/azure/foundry/foundry-models/how-to/use-foundry-models-microsoft-decision)
and [TypeSafe System One API](https://docs.typesafe.ai/api/system-one).

---

## Usage

### Single async call

```python
import asyncio
from ehs_llm_client import LLM

async def main():
    llm = LLM("prod", config_file_path="llmconfig.cfg")
    messages = [{"role": "user", "content": "Say hello in JSON"}]

    response, in_tokens, out_tokens = await llm.get_response_async(
        messages, schema="return json"
    )

    print(response)

asyncio.run(main())
```

One `LLM` can be used from several event loops in turn, for example a Celery task
that calls `asyncio.run()` for each group of work. Since 0.1.13, `llm.client` is
replaced with a new provider client the first time it is used from a new event
loop. Before that, calls in a later loop reused connections opened by the closed
one; they failed, and the OpenAI SDK retried them ("Retrying request to
/responses"). Assigning `llm.client` yourself still works; the next event loop
to use it adopts it.

### Connections and concurrency

Hosts such as Azure App Service give each instance a limited number of outbound
(SNAT) ports per destination, about 128. Every open connection uses one. When
they run out, new connections time out (`httpx.ConnectTimeout`, then
`openai.APITimeoutError`) while existing ones keep working. Three settings,
read from the environment, control how many connections the client opens:

| Variable | Default | Effect |
|---|---|---|
| `LLM_MAX_CONCURRENCY` | `0` (no limit) | Most `get_response_async` requests in flight at once, shared by every `LLM` in the process (per event loop). Further calls wait for a slot. Waiting does not count toward `timeout`, and a call holds no slot while it sleeps between retries. Since 0.1.15. |
| `LLM_HTTP2` | off | `true` sends OpenAI and Azure OpenAI requests over HTTP/2, so concurrent requests share connections instead of opening one each. Servers without HTTP/2 are spoken to over HTTP/1.1. Since 0.1.15. |
| `LLM_KEEPALIVE_EXPIRY` | `60` | Seconds an idle connection is kept for reuse (the OpenAI SDK default is 5), so calls that come in groups reuse connections instead of opening new ones. Must be more than 0 and less than 240: Azure drops outbound connections idle for longer without telling either end. Since 0.1.14. |

The ports a process can use at once are about `LLM_MAX_CONCURRENCY` over
HTTP/1.1, and far fewer over HTTP/2. Several processes on one instance share
its ports, so size the limit as the instance's budget divided by the number of
processes that make calls. A lower limit means calls in a large burst wait for
each other; a burst that is never larger than the limit is not slowed.

HTTP/2 trade-offs: a connection that drops or loses packets affects every
request on it at once, and httpx's HTTP/2 support has seen less use than its
HTTP/1.1 support. Turn it on in a test environment first.

Other SDK defaults are unchanged: connect timeout 5 s, read timeout 600 s, at
most 1000 connections per client. Batch file and job calls are not limited by
`LLM_MAX_CONCURRENCY`.

Blocked, failed or incomplete OpenAI/Azure responses raise
`ehs_llm_client.exceptions.LLMResponseError` (a `RuntimeError`) instead of returning
partial or empty text as success. Content-filter errors are not retried with the same prompt.

### Reasoning effort

The reasoning controls from version 0.1.10 are supported:

```ini
[reasoning]
provider = openai
model = gpt-6-astra
default_reasoning_effort = medium
```

Set `default_reasoning_effort` in a config section/dictionary, or set
`LLM_DEFAULT_REASONING_EFFORT` when using environment configuration. Override it
for a single call with `reasoning_effort`:

```python
response, in_tokens, out_tokens = await llm.get_response_async(
    messages, schema="return json", reasoning_effort="high"
)
```

`call_async` also accepts `reasoning_effort`. OpenAI uses the model support check
from 0.1.10. Azure forwards an explicitly configured effort because deployment
names can be arbitrary; only set it for a reasoning-capable deployment. Supported
effort values depend on the model. If you omit the setting, the API's default applies.
See [OpenAI reasoning guidance](https://developers.openai.com/api/docs/guides/reasoning).
The setting applies to single calls; batch behavior remains as in 0.1.10.

The 0.1.10 method signatures are preserved, including positional reasoning,
retry and timeout arguments. Prefer keyword arguments for clarity.

OpenAI/Azure JSON mode (`schema="return json"`) requires string message contents,
as in 0.1.10. Regular calls and schema-based output retain support for content arrays.

### Native typed decisions

Use the same questions with either a Foundry `systemone` section or a TypeSafe
section. State can be text, a JSON object or an array. Instructions and criteria
descriptions may also contain JSON structures.

```python
import asyncio
from ehs_llm_client import LLM, ChoiceAnswer

async def main():
    async with LLM("typesafe_jev", config_file_path="llmconfig.cfg") as model:
        result = await model.decide_async(
            state="A worker suffered a minor cut and received first aid.",
            questions={
                "category": {
                    "type": "choice",
                    "instructions": "Classify the reported incident.",
                    "criteria": {
                        "injury": "An injury occurred.",
                        "near_miss": "An event occurred without injury.",
                        "unknown": "There is insufficient information.",
                    },
                },
                "injured": {"type": "noul", "instructions": "Was anyone injured?"},
                "severity": {
                    "type": "score",
                    "instructions": "Rate the reported outcome.",
                    "criteria": ["No injury", "Minor injury", "Serious injury"],
                },
            },
            max_retries=3,
            timeout=30,
        )
        answer = result.answers["category"]
        if isinstance(answer, ChoiceAnswer):
            print(answer.choice, answer.probabilities, answer.confidence)
        print(result.model, result.usage.input_tokens, result.usage.output_tokens)

asyncio.run(main())
```

`DecisionResult` contains typed `ChoiceAnswer`, `NoulAnswer` and `ScoreAnswer`
objects, a resolved `model`, `DecisionUsage`, an optional `request_id`, and the
complete provider JSON in `raw_response`. Choice/score probabilities, confidence
and score legends are preserved when supplied; absent optional fields are `None`.
Noul is a probability from 0 to 1, not a boolean. Score is a number between zero
and the last criterion's index and can fall between levels. The library does not
choose thresholds, turn uncertainty into labels, or translate answers for EIDA.

Input and answer shapes, allowed labels, numeric ranges and token usage are
validated. Malformed successful responses raise `LLMResponseError`. HTTP failures
raise `LLMAPIError` (a subclass) with `status_code` and optional `request_id`;
error messages exclude provider response bodies. Inspect `raw_response` carefully
before logging it because provider extensions can contain sensitive information.

Decision calls retry transport failures, timeouts and HTTP 408, 429, 500, 502,
503, 504 and 529, honoring `Retry-After` and `retry-after-ms` when valid. Other HTTP
failures and invalid answers are not retried. As with the existing library API,
`max_retries` means **total attempts**, including the first request; `1` disables
retries. `timeout` is the limit for each attempt, so retries and backoff can make
the complete call longer. Cancellation propagates to the caller.

System One clients reject generative calls, provider batch operations,
`model_batch`, and `default_reasoning_effort`. Use concurrent `decide_async()`
calls for application-level parallelism. Generative clients reject `decide_async()`.

All providers support `async with LLM(...)` and `await llm.aclose()` to release
their clients. Keep a client open across repeated calls to reuse connections.

---

### Batch request

```python
batch_line = llm.create_batch_request(
    custom_id="test1",
    messages=[{"role": "user", "content": "Give me a JSON object"}],
    schema="return json"
)

batch_id = await llm.run_batch_process([batch_line], submission_id="batch123")
status = await llm.get_batch_status(batch_id)
print(status)
```

`create_batch_request` returns a dictionary. `run_batch_process` accepts these
dictionaries or JSONL strings that you already serialized. When `model_batch` is
omitted, OpenAI/Azure batches use `model`.

For Azure, set `model_batch` to a deployment that supports batch processing
(`GlobalBatch` or `DataZoneBatch`). Model and region availability can differ from
single-call deployments; see [Azure batch documentation](https://learn.microsoft.com/en-us/azure/foundry/openai/how-to/batch).
Azure supports submission, status checking, cancellation and result downloads
through the same client methods.

---

## API Key Management

Keep API keys **out of code**. Recommended environment variables:

| Provider | Standard env var | Existing env vars (still supported) |
|----------|------------------|-----------------------------------|
| OpenAI | `OPENAI_API_KEY` | `openaiAPIkey`, `openaiAPIKey` |
| Azure OpenAI | `AZURE_OPENAI_API_KEY` | `azure_openaiAPIkey`, `azure_openaiAPIKey` |
| Google Gemini | `GOOGLE_API_KEY` | `googleAPIkey`, `googleAPIKey` |
| Microsoft Foundry | `AZURE_FOUNDRY_API_KEY` | `azure_foundryAPIkey`, `azure_foundryAPIKey` |
| TypeSafe | `TYPESAFE_API_KEY` | `typesafeAPIkey`, `typesafeAPIKey` |

The 0.1.10 `APIKey` spelling takes precedence, followed by the older `APIkey`
spelling and then the standard variable. Azure credentials must come from your Azure resource.

You can also inject an in-memory key using the new keyword-only argument
`LLM(..., api_key=secret_value)`. This takes precedence over all environment
variables and works for existing providers too. Retrieve Key Vault secrets in
your application or notebook and pass the value; this package has no dependency
on a particular vault, Azure identity configuration or DataTeam filesystem path.
Do not store keys in configuration files or commit them to source control.

---

## Development

```bash
# Create a virtual environment
python -m venv .venv
source .venv/bin/activate  # macOS/Linux
.venv\Scripts\activate     # Windows

# Install dependencies
pip install -e .
```

Run the regression checks (simulated HTTP responses; no API credentials needed):

```bash
python -m unittest discover -s tests -v
```

---

## License

MIT License © 2026 Your Name  
