llm-switchboard 0.1.0

A provider-agnostic Python interface for routing generation calls across LiteLLM, OpenAI, Ollama, and Watsonx

LLM Switchboard

LLM Switchboard is a small Python library for calling different model providers through one consistent interface. It supports synchronous and asynchronous generation, tool calls, structured output validation, provider-specific parameter mapping, and configurable logging.

Installation

pip install llm-switchboard

For local development:

git clone [e-mail]:skillberry-ai/llm-switchboard.git
cd llm-switchboard
pip install -e ".[dev,all]"

LiteLLM is installed by default. Additional provider extras are available when you need direct SDK support:

pip install "llm-switchboard[openai]"
pip install "llm-switchboard[watsonx]"
pip install "llm-switchboard[all]"

Quick Start

from llm_switchboard.llm import get_llm

LiteLLMClient = get_llm("litellm")
client = LiteLLMClient(model_name="openai/gpt-4o-mini")

response = client.generate("Write a short explanation of retrieval augmented generation.")
print(response)

You can also select a provider from environment variables:

export LLM_PROVIDER=litellm
export MODEL_NAME=openai/gpt-4o-mini
export OPENAI_API_KEY=...

from llm_switchboard.llm import get_llm

client = get_llm("auto_from_env")()
print(client.generate("Summarize the role of a model gateway."))

If LLM_PROVIDER is not set, auto_from_env defaults to litellm. Providers that need a model name still require MODEL_NAME.

Providers

LiteLLM

LiteLLM is the default provider and works with any LiteLLM-supported model string.

from llm_switchboard.llm import get_llm

client = get_llm("litellm")(model_name="openai/gpt-4o-mini")
response = client.generate(
    [{"role": "user", "content": "Return three release checklist items."}],
    temperature=0.2,
)

Registered names:

• litellm
• litellm.output_val

OpenAI and Azure OpenAI

Install the direct OpenAI SDK extra when you want to use OpenAI clients without the LiteLLM adapter.

pip install "llm-switchboard[openai]"

from llm_switchboard.llm import get_llm

client = get_llm("openai.sync")(api_key="...")
response = client.generate("Explain semantic versioning.", model="gpt-4o-mini")

Registered names:

• openai.sync
• openai.async
• openai.sync.output_val
• openai.async.output_val
• azure_openai.sync
• azure_openai.async
• azure_openai.sync.output_val
• azure_openai.async.output_val

Ollama

Ollama uses the LiteLLM dependency that is installed by default.

from llm_switchboard.llm import get_llm

client = get_llm("litellm.ollama")(
    model_name="llama3.1",
    api_base="http://localhost:11434",
)
print(client.generate("Give me one practical use for local models."))

Registered names:

• litellm.ollama
• litellm.ollama.output_val

Watsonx

Watsonx can be used through LiteLLM or the native SDK.

pip install "llm-switchboard[watsonx]"

from llm_switchboard.llm import get_llm

client = get_llm("litellm.watsonx")(
    model_name="meta-llama/llama-3-3-70b-instruct",
    api_key="...",
    project_id="...",
)

Registered names:

• litellm.watsonx
• litellm.watsonx.output_val
• watsonx
• watsonx.output_val

Structured Output

Use an output-validation provider when you want a Pydantic model, JSON schema, or built-in type back from the response.

from pydantic import BaseModel

from llm_switchboard.llm import get_llm


class Task(BaseModel):
    title: str
    priority: str


client = get_llm("litellm.output_val")(model_name="openai/gpt-4o-mini")
task = client.generate(
    "Create one task for preparing a package release.",
    output_schema=Task,
)
print(task.title, task.priority)

Generation Arguments

GenerationArgs provides provider-agnostic parameters such as max_tokens, temperature, top_p, seed, timeout, stop_sequences, and stream.

from llm_switchboard.llm import get_llm
from llm_switchboard.llm.types import GenerationArgs

client = get_llm("litellm")(model_name="openai/gpt-4o-mini")
args = GenerationArgs(max_tokens=120, temperature=0.3, seed=7)

response = client.generate("Draft a concise changelog entry.", generation_args=args)

Each provider maps supported arguments to its native SDK shape and ignores unsupported values with a warning where appropriate.

Logging

from llm_switchboard import configure_logging

configure_logging(level="INFO")

Environment variables:

• LLM_SWITCHBOARD_LOG_LEVEL
• LLM_SWITCHBOARD_LOG_DIR

Sensitive keys such as API keys, tokens, credentials, and authorization headers are masked before logging.

Examples

Example scripts live in llm_examples/:

• litellm_ollama_example.py
• azure_openai_example.py
• litellm_watsonx_example.py
• ibm_watsonx_ai_example.py
• litellm_rits_example.py
• litellm_ibm_example.py

Run an example from the repository root:

python llm_examples/litellm_ollama_example.py

Enterprise/Internal Integrations

RITS and IBM ETE LiteLLM adapters are included for internal IBM services.

RITS:

• Provider names: litellm.rits, litellm.rits.output_val
• Environment variables: RITS_API_KEY, RITS_API_URL

IBM ETE LiteLLM:

• Provider names: litellm.ibm, litellm.ibm.output_val
• Environment variables: IBM_THIRD_PARTY_API_KEY, IBM_LITELLM_API_BASE

Development

pip install -e ".[dev,all]"
make lint
make coverage
make build

Release builds are created with:

python -m build
twine check dist/*

License

Apache-2.0