Models

Sonika AI Toolkit supports five LLM providers. All implement the ILanguageModel interface with predict(), invoke(), and stream_response() methods.

OpenAI

from sonika_ai_toolkit import OpenAILanguageModel

llm = OpenAILanguageModel(
    api_key="sk-...",
    model_name="gpt-4o-mini",  # default
    temperature=0.7,
)

Supported models: gpt-4o, gpt-4o-mini, gpt-4-turbo, o1, o1-mini, etc.

Google Gemini

from sonika_ai_toolkit import GeminiLanguageModel

llm = GeminiLanguageModel(
    api_key="...",
    model_name="gemini-2.5-flash",
    temperature=0.7,
)

Thinking models

Gemini thinking models (gemini-2.5-*, *-thinking, *thinking-exp*) require temperature=1.0 — this is automatically overridden with a warning.

Thinking model behavior: When include_thoughts=True, response.content is a list containing thinking and text blocks. The toolkit handles this automatically.

DeepSeek

from sonika_ai_toolkit import DeepSeekLanguageModel

llm = DeepSeekLanguageModel(
    api_key="...",
    model_name="deepseek-chat",  # or "deepseek-reasoner"
)

DeepSeek Reasoner

deepseek-reasoner (and models with r1 in the name) uses a custom _DeepSeekReasonerChatModel that captures reasoning_content. It does not support tool calling — a ValueError is raised if tools are provided.

Anthropic (Claude)

from sonika_ai_toolkit import AnthropicLanguageModel

llm = AnthropicLanguageModel(
    api_key="sk-ant-...",
    model_name="claude-haiku-4-5",  # default
    temperature=0.7,
    max_tokens=4096,                # Anthropic requires max_tokens
)

Supported models: claude-haiku-4-5, claude-sonnet-4-6, claude-opus-4-8, etc.

Extended thinking

Pass thinking_budget=N to enable extended thinking. This forces temperature=1.0 (with a warning) and raises max_tokens above the budget. As with Gemini, thinking responses return content as a list of blocks; the toolkit strips the thinking blocks automatically.

Amazon Bedrock

from sonika_ai_toolkit import BedrockLanguageModel

llm = BedrockLanguageModel(
    api_key="...",
    model_name="amazon.nova-micro-v1:0",
    region="us-east-1",
)

The AWS_BEARER_TOKEN_BEDROCK env var is set automatically during initialization.

ILanguageModel Interface

All models implement:

from sonika_ai_toolkit import ILanguageModel

class ILanguageModel(ABC):
    model: BaseChatModel  # underlying LangChain model

    def predict(self, prompt: str) -> str: ...
    def invoke(self, prompt: str) -> ResponseModel: ...
    def stream_response(self, prompt: str) -> Iterator: ...

Type-hint against ILanguageModel for provider-agnostic code:

def classify(llm: ILanguageModel, text: str):
    return llm.predict(f"Classify: {text}")

Top-Level Imports

from sonika_ai_toolkit import (
    OpenAILanguageModel,
    GeminiLanguageModel,
    DeepSeekLanguageModel,
    BedrockLanguageModel,
    AnthropicLanguageModel,
    ILanguageModel,
)