Language Models API

LanguageModel

Bases: SynalinksSaveable

A language model API wrapper.

A language model is a type of AI model designed to generate, and interpret human language. It is trained on large amounts of text data to learn patterns and structures in language. Language models can perform various tasks such as text generation, translation, summarization, and answering questions.

We support providers that implement constrained structured output like OpenAI, Azure, Ollama or Mistral. In addition we support providers that otherwise allow to constrain the use of a specific tool like Groq or Anthropic.

For the complete list of models, please refer to the providers documentation.

Using OpenAI models

import synalinks
import os

os.environ["OPENAI_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="openai/gpt-4o-mini",
)

Using Groq models

import synalinks
import os

os.environ["GROQ_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="groq/llama3-8b-8192",
)

Using Anthropic models

import synalinks
import os

os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="anthropic/claude-3-sonnet-20240229",
)

Using Mistral models

import synalinks
import os

os.environ["MISTRAL_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="mistral/codestral-latest",
)

Using Ollama models

import synalinks
import os

language_model = synalinks.LanguageModel(
    model="ollama/mistral",
)

Using Azure OpenAI models

import synalinks
import os

os.environ["AZURE_API_KEY"] = "your-api-key"
os.environ["AZURE_API_BASE"] = "your-api-key"
os.environ["AZURE_API_VERSION"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="azure/<your_deployment_name>",
)

To cascade models in case there is anything wrong with the model provider (hence making your pipelines more robust). Use the fallback argument like in this example:

import synalinks
import os

os.environ["OPENAI_API_KEY"] = "your-api-key"
os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

language_model = synalinks.LanguageModel(
    model="anthropic/claude-3-sonnet-20240229",
    fallback=synalinks.LanguageModel(
        model="openai/gpt-4o-mini",
    )
)

Note: Obviously, use an .env file and .gitignore to avoid putting your API keys in the code or a config file that can lead to leackage when pushing it into repositories.

Parameters:

Name	Type	Description	Default
model	str	The model to use.	None
api_base	str	Optional. The endpoint to use.	None
retry	int	Optional. The number of retry (default to 5).	5
fallback	LanguageModel	Optional. The language model to fallback if anything is wrong.	None

Source code in synalinks/src/language_models/language_model.py

@synalinks_export(
    [
        "synalinks.LanguageModel",
        "synalinks.language_models.LanguageModel",
    ]
)
class LanguageModel(SynalinksSaveable):
    """A language model API wrapper.

    A language model is a type of AI model designed to generate, and interpret human
    language. It is trained on large amounts of text data to learn patterns and
    structures in language. Language models can perform various tasks such as text
    generation, translation, summarization, and answering questions.

    We support providers that implement *constrained structured output*
    like OpenAI, Azure, Ollama or Mistral. In addition we support providers that otherwise
    allow to constrain the use of a specific tool like Groq or Anthropic.

    For the complete list of models, please refer to the providers documentation.

    **Using OpenAI models**

    ```python
    import synalinks
    import os

    os.environ["OPENAI_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="openai/gpt-4o-mini",
    )
    ```

    **Using Groq models**

    ```python
    import synalinks
    import os

    os.environ["GROQ_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="groq/llama3-8b-8192",
    )
    ```

    **Using Anthropic models**

    ```python
    import synalinks
    import os

    os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="anthropic/claude-3-sonnet-20240229",
    )
    ```

    **Using Mistral models**

    ```python
    import synalinks
    import os

    os.environ["MISTRAL_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="mistral/codestral-latest",
    )
    ```

    **Using Ollama models**

    ```python
    import synalinks
    import os

    language_model = synalinks.LanguageModel(
        model="ollama/mistral",
    )
    ```

    **Using Azure OpenAI models**

    ```python
    import synalinks
    import os

    os.environ["AZURE_API_KEY"] = "your-api-key"
    os.environ["AZURE_API_BASE"] = "your-api-key"
    os.environ["AZURE_API_VERSION"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="azure/<your_deployment_name>",
    )
    ```

    To cascade models in case there is anything wrong with
    the model provider (hence making your pipelines more robust).
    Use the `fallback` argument like in this example:

    ```python
    import synalinks
    import os

    os.environ["OPENAI_API_KEY"] = "your-api-key"
    os.environ["ANTHROPIC_API_KEY"] = "your-api-key"

    language_model = synalinks.LanguageModel(
        model="anthropic/claude-3-sonnet-20240229",
        fallback=synalinks.LanguageModel(
            model="openai/gpt-4o-mini",
        )
    )
    ```

    **Note**: Obviously, use an `.env` file and `.gitignore` to avoid
    putting your API keys in the code or a config file that can lead to
    leackage when pushing it into repositories.

    Args:
        model (str): The model to use.
        api_base (str): Optional. The endpoint to use.
        retry (int): Optional. The number of retry (default to 5).
        fallback (LanguageModel): Optional. The language model to fallback
            if anything is wrong.
    """

    def __init__(
        self,
        model=None,
        api_base=None,
        retry=5,
        fallback=None,
    ):
        if model is None:
            raise ValueError("You need to set the `model` argument for any LanguageModel")
        model_provider = model.split("/")[0]
        if model_provider == "ollama":
            # Switch from `ollama` to `ollama_chat`
            # because it have better performance due to the chat prompts
            model = model.replace("ollama", "ollama_chat")
        self.model = model
        self.fallback = fallback
        if self.model.startswith("ollama") and not api_base:
            self.api_base = "http://localhost:11434"
        else:
            self.api_base = api_base
        self.retry = retry

    async def __call__(self, messages, schema=None, streaming=False, **kwargs):
        """
        Call method to generate a response using the language model.

        Args:
            messages (dict): A formatted dict of chat messages.
            schema (dict): The target JSON schema for structed output (optional).
                If None, output a ChatMessage-like answer.
            streaming (bool): Enable streaming (optional). Default to False.
                Can be enabled only if schema is None.
            functions (list): A list of Python functions for the LM to use.
            **kwargs (keyword arguments): The additional keywords arguments
                forwarded to the LM call.
        Returns:
            (dict): The generated structured response.
        """
        maybe_initialize_telemetry()
        formatted_messages = messages.get_json().get("messages", [])
        json_instance = {}
        input_kwargs = copy.deepcopy(kwargs)
        if schema:
            if self.model.startswith("groq"):
                # Use a tool created on the fly for groq
                kwargs.update(
                    {
                        "tools": [
                            {
                                "function": {
                                    "name": "structured_output",
                                    "description": "Generate a valid JSON output",
                                    "parameters": schema.get("properties"),
                                },
                                "type": "function",
                            }
                        ],
                        "tool_choice": {
                            "type": "function",
                            "function": {"name": "structured_output"},
                        },
                    }
                )
            elif self.model.startswith("anthropic"):
                # Use a tool created on the fly for anthropic
                kwargs.update(
                    {
                        "tools": [
                            {
                                "name": "structured_output",
                                "description": "Generate a valid JSON output",
                                "input_schema": {
                                    "type": "object",
                                    "properties": schema.get("properties"),
                                    "required": schema.get("required"),
                                },
                            }
                        ],
                        "tool_choice": {
                            "type": "tool",
                            "name": "structured_output",
                        },
                    }
                )
            elif self.model.startswith("ollama") or self.model.startswith("mistral"):
                # Use constrained structured output for ollama/mistral
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": {"schema": schema},
                            "strict": True,
                        },
                    }
                )
            elif self.model.startswith("openai") or self.model.startswith("azure"):
                # Use constrained structured output for openai
                # OpenAI require the field  "additionalProperties"
                kwargs.update(
                    {
                        "response_format": {
                            "type": "json_schema",
                            "json_schema": {
                                "name": "structured_output",
                                "strict": True,
                                "schema": schema,
                            },
                        }
                    }
                )
            else:
                provider = self.model.split("/")[0]
                raise ValueError(
                    f"LM provider '{provider}' not supported yet, please ensure that"
                    " they support constrained structured output and fill an issue."
                )

        if self.api_base:
            kwargs.update(
                {
                    "api_base": self.api_base,
                }
            )
        if streaming and schema:
            streaming = False
        if streaming:
            kwargs.update({"stream": True})
        for i in range(self.retry):
            try:
                response_str = ""
                response = litellm.completion(
                    model=self.model,
                    messages=formatted_messages,
                    caching=False,
                    **kwargs,
                )
                if streaming:
                    return StreamingIterator(response)
                if (
                    self.model.startswith("groq") or self.model.startswith("anthropic")
                ) and schema:
                    response_str = response["choices"][0]["message"]["tool_calls"][0][
                        "function"
                    ]["arguments"]
                else:
                    response_str = response["choices"][0]["message"]["content"].strip()
                if schema:
                    json_instance = json.loads(response_str)
                else:
                    json_instance = {"role": ChatRole.ASSISTANT, "content": response_str}
                return json_instance
            except Exception as e:
                warnings.warn(f"Error occured while trying to call {self}: " + str(e))
                capture_exception(e)
        if self.fallback:
            return self.fallback(
                messages,
                schema=schema,
                streaming=streaming,
                **input_kwargs,
            )
        else:
            return None

    def _obj_type(self):
        return "LanguageModel"

    def get_config(self):
        config = {
            "model": self.model,
            "api_base": self.api_base,
            "retry": self.retry,
        }
        if self.fallback:
            fallback_config = {
                "fallback": serialization_lib.serialize_synalinks_object(
                    self.fallback,
                )
            }
            return {**fallback_config, **config}
        else:
            return config

    @classmethod
    def from_config(cls, config):
        if "fallback" in config:
            fallback = serialization_lib.deserialize_synalinks_object(
                config.pop("fallback")
            )
            return cls(fallback=fallback, **config)
        else:
            return cls(**config)

    def __repr__(self):
        api_base = f" api_base={self.api_base}" if self.api_base else ""
        return f"<LanguageModel model={self.model}{api_base}>"

__call__(messages, schema=None, streaming=False, **kwargs) async

Call method to generate a response using the language model.

Parameters:

Name	Type	Description	Default
messages	dict	A formatted dict of chat messages.	required
schema	dict	The target JSON schema for structed output (optional). If None, output a ChatMessage-like answer.	None
streaming	bool	Enable streaming (optional). Default to False. Can be enabled only if schema is None.	False
functions	list	A list of Python functions for the LM to use.	required
**kwargs	keyword arguments	The additional keywords arguments forwarded to the LM call.	{}

Returns: (dict): The generated structured response.

Source code in synalinks/src/language_models/language_model.py

async def __call__(self, messages, schema=None, streaming=False, **kwargs):
    """
    Call method to generate a response using the language model.

    Args:
        messages (dict): A formatted dict of chat messages.
        schema (dict): The target JSON schema for structed output (optional).
            If None, output a ChatMessage-like answer.
        streaming (bool): Enable streaming (optional). Default to False.
            Can be enabled only if schema is None.
        functions (list): A list of Python functions for the LM to use.
        **kwargs (keyword arguments): The additional keywords arguments
            forwarded to the LM call.
    Returns:
        (dict): The generated structured response.
    """
    maybe_initialize_telemetry()
    formatted_messages = messages.get_json().get("messages", [])
    json_instance = {}
    input_kwargs = copy.deepcopy(kwargs)
    if schema:
        if self.model.startswith("groq"):
            # Use a tool created on the fly for groq
            kwargs.update(
                {
                    "tools": [
                        {
                            "function": {
                                "name": "structured_output",
                                "description": "Generate a valid JSON output",
                                "parameters": schema.get("properties"),
                            },
                            "type": "function",
                        }
                    ],
                    "tool_choice": {
                        "type": "function",
                        "function": {"name": "structured_output"},
                    },
                }
            )
        elif self.model.startswith("anthropic"):
            # Use a tool created on the fly for anthropic
            kwargs.update(
                {
                    "tools": [
                        {
                            "name": "structured_output",
                            "description": "Generate a valid JSON output",
                            "input_schema": {
                                "type": "object",
                                "properties": schema.get("properties"),
                                "required": schema.get("required"),
                            },
                        }
                    ],
                    "tool_choice": {
                        "type": "tool",
                        "name": "structured_output",
                    },
                }
            )
        elif self.model.startswith("ollama") or self.model.startswith("mistral"):
            # Use constrained structured output for ollama/mistral
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": {"schema": schema},
                        "strict": True,
                    },
                }
            )
        elif self.model.startswith("openai") or self.model.startswith("azure"):
            # Use constrained structured output for openai
            # OpenAI require the field  "additionalProperties"
            kwargs.update(
                {
                    "response_format": {
                        "type": "json_schema",
                        "json_schema": {
                            "name": "structured_output",
                            "strict": True,
                            "schema": schema,
                        },
                    }
                }
            )
        else:
            provider = self.model.split("/")[0]
            raise ValueError(
                f"LM provider '{provider}' not supported yet, please ensure that"
                " they support constrained structured output and fill an issue."
            )

    if self.api_base:
        kwargs.update(
            {
                "api_base": self.api_base,
            }
        )
    if streaming and schema:
        streaming = False
    if streaming:
        kwargs.update({"stream": True})
    for i in range(self.retry):
        try:
            response_str = ""
            response = litellm.completion(
                model=self.model,
                messages=formatted_messages,
                caching=False,
                **kwargs,
            )
            if streaming:
                return StreamingIterator(response)
            if (
                self.model.startswith("groq") or self.model.startswith("anthropic")
            ) and schema:
                response_str = response["choices"][0]["message"]["tool_calls"][0][
                    "function"
                ]["arguments"]
            else:
                response_str = response["choices"][0]["message"]["content"].strip()
            if schema:
                json_instance = json.loads(response_str)
            else:
                json_instance = {"role": ChatRole.ASSISTANT, "content": response_str}
            return json_instance
        except Exception as e:
            warnings.warn(f"Error occured while trying to call {self}: " + str(e))
            capture_exception(e)
    if self.fallback:
        return self.fallback(
            messages,
            schema=schema,
            streaming=streaming,
            **input_kwargs,
        )
    else:
        return None