r2rpy/py/shared/abstractions/llm.py

"""Abstractions for the LLM model."""

import json
from enum import Enum
from typing import TYPE_CHECKING, Any, ClassVar, Optional

from openai.types.chat import ChatCompletionChunk
from pydantic import BaseModel, Field

from .base import R2RSerializable

if TYPE_CHECKING:
    from .search import AggregateSearchResult

from typing_extensions import Literal


class Function(BaseModel):
    arguments: str
    """
    The arguments to call the function with, as generated by the model in JSON
    format. Note that the model does not always generate valid JSON, and may
    hallucinate parameters not defined by your function schema. Validate the
    arguments in your code before calling your function.
    """

    name: str
    """The name of the function to call."""


class ChatCompletionMessageToolCall(BaseModel):
    id: str
    """The ID of the tool call."""

    function: Function
    """The function that the model called."""

    type: Literal["function"]
    """The type of the tool. Currently, only `function` is supported."""


class FunctionCall(BaseModel):
    arguments: str
    """
    The arguments to call the function with, as generated by the model in JSON
    format. Note that the model does not always generate valid JSON, and may
    hallucinate parameters not defined by your function schema. Validate the
    arguments in your code before calling your function.
    """

    name: str
    """The name of the function to call."""


class ChatCompletionMessage(BaseModel):
    content: Optional[str] = None
    """The contents of the message."""

    refusal: Optional[str] = None
    """The refusal message generated by the model."""

    role: Literal["assistant"]
    """The role of the author of this message."""

    # audio: Optional[ChatCompletionAudio] = None
    """
    If the audio output modality is requested, this object contains data about the
    audio response from the model.
    [Learn more](https://platform.openai.com/docs/guides/audio).
    """

    function_call: Optional[FunctionCall] = None
    """Deprecated and replaced by `tool_calls`.

    The name and arguments of a function that should be called, as generated by the
    model.
    """

    tool_calls: Optional[list[ChatCompletionMessageToolCall]] = None
    """The tool calls generated by the model, such as function calls."""

    structured_content: Optional[list[dict]] = None


class Choice(BaseModel):
    finish_reason: Literal[
        "stop",
        "length",
        "tool_calls",
        "content_filter",
        "function_call",
        "max_tokens",
    ]
    """The reason the model stopped generating tokens.

    This will be `stop` if the model hit a natural stop point or a provided stop
    sequence, `length` if the maximum number of tokens specified in the request was
    reached, `content_filter` if content was omitted due to a flag from our content
    filters, `tool_calls` if the model called a tool, or `function_call`
    (deprecated) if the model called a function.
    """

    index: int
    """The index of the choice in the list of choices."""

    # logprobs: Optional[ChoiceLogprobs] = None
    """Log probability information for the choice."""

    message: ChatCompletionMessage
    """A chat completion message generated by the model."""


class LLMChatCompletion(BaseModel):
    id: str
    """A unique identifier for the chat completion."""

    choices: list[Choice]
    """A list of chat completion choices.

    Can be more than one if `n` is greater than 1.
    """

    created: int
    """The Unix timestamp (in seconds) of when the chat completion was created."""

    model: str
    """The model used for the chat completion."""

    object: Literal["chat.completion"]
    """The object type, which is always `chat.completion`."""

    service_tier: Optional[Literal["scale", "default"]] = None
    """The service tier used for processing the request."""

    system_fingerprint: Optional[str] = None
    """This fingerprint represents the backend configuration that the model runs with.

    Can be used in conjunction with the `seed` request parameter to understand when
    backend changes have been made that might impact determinism.
    """

    usage: Optional[Any] = None
    """Usage statistics for the completion request."""


LLMChatCompletionChunk = ChatCompletionChunk


class RAGCompletion:
    completion: LLMChatCompletion
    search_results: "AggregateSearchResult"

    def __init__(
        self,
        completion: LLMChatCompletion,
        search_results: "AggregateSearchResult",
    ):
        self.completion = completion
        self.search_results = search_results


class GenerationConfig(R2RSerializable):
    _defaults: ClassVar[dict] = {
        "model": "openai/gpt-4.1-mini",
        "temperature": 0.1,
        "top_p": 1.0,
        "max_tokens_to_sample": 102400,
        "stream": False,
        "functions": None,
        "tools": None,
        "add_generation_kwargs": None,
        "api_base": None,
        "response_format": None,
        "extended_thinking": False,
        "thinking_budget": None,
        "reasoning_effort": None,
    }

    model: Optional[str] = Field(
        default_factory=lambda: GenerationConfig._defaults["model"]
    )
    temperature: float = Field(
        default_factory=lambda: GenerationConfig._defaults["temperature"]
    )
    top_p: Optional[float] = Field(
        default_factory=lambda: GenerationConfig._defaults["top_p"],
    )
    max_tokens_to_sample: int = Field(
        default_factory=lambda: GenerationConfig._defaults[
            "max_tokens_to_sample"
        ],
    )
    stream: bool = Field(
        default_factory=lambda: GenerationConfig._defaults["stream"]
    )
    functions: Optional[list[dict]] = Field(
        default_factory=lambda: GenerationConfig._defaults["functions"]
    )
    tools: Optional[list[dict]] = Field(
        default_factory=lambda: GenerationConfig._defaults["tools"]
    )
    add_generation_kwargs: Optional[dict] = Field(
        default_factory=lambda: GenerationConfig._defaults[
            "add_generation_kwargs"
        ],
    )
    api_base: Optional[str] = Field(
        default_factory=lambda: GenerationConfig._defaults["api_base"],
    )
    response_format: Optional[dict | BaseModel] = None
    extended_thinking: bool = Field(
        default=False,
        description="Flag to enable extended thinking mode (for Anthropic providers)",
    )
    thinking_budget: Optional[int] = Field(
        default=None,
        description=(
            "Token budget for internal reasoning when extended thinking mode is enabled. "
            "Must be less than max_tokens_to_sample."
        ),
    )
    reasoning_effort: Optional[str] = Field(
        default=None,
        description=(
            "Effort level for internal reasoning when extended thinking mode is enabled, `low`, `medium`, or `high`."
            "Only applicable to OpenAI providers."
        ),
    )

    @classmethod
    def set_default(cls, **kwargs):
        for key, value in kwargs.items():
            if key in cls._defaults:
                cls._defaults[key] = value
            else:
                raise AttributeError(
                    f"No default attribute '{key}' in GenerationConfig"
                )

    def __init__(self, **data):
        # Handle max_tokens mapping to max_tokens_to_sample
        if "max_tokens" in data:
            # Only set max_tokens_to_sample if it's not already provided
            if "max_tokens_to_sample" not in data:
                data["max_tokens_to_sample"] = data.pop("max_tokens")
            else:
                # If both are provided, max_tokens_to_sample takes precedence
                data.pop("max_tokens")

        if (
            "response_format" in data
            and isinstance(data["response_format"], type)
            and issubclass(data["response_format"], BaseModel)
        ):
            model_class = data["response_format"]
            data["response_format"] = {
                "type": "json_schema",
                "json_schema": {
                    "name": model_class.__name__,
                    "schema": model_class.model_json_schema(),
                },
            }

        model = data.pop("model", None)
        if model is not None:
            super().__init__(model=model, **data)
        else:
            super().__init__(**data)

    def __str__(self):
        return json.dumps(self.to_dict())

    class Config:
        populate_by_name = True
        json_schema_extra = {
            "example": {
                "model": "openai/gpt-4.1",
                "temperature": 0.1,
                "top_p": 1.0,
                "max_tokens_to_sample": 1024,
                "stream": False,
                "functions": None,
                "tools": None,
                "add_generation_kwargs": None,
                "api_base": None,
            }
        }


class MessageType(Enum):
    SYSTEM = "system"
    USER = "user"
    ASSISTANT = "assistant"
    FUNCTION = "function"
    TOOL = "tool"

    def __str__(self):
        return self.value


class Message(R2RSerializable):
    role: MessageType | str
    content: Optional[Any] = None
    name: Optional[str] = None
    function_call: Optional[dict[str, Any]] = None
    tool_calls: Optional[list[dict[str, Any]]] = None
    tool_call_id: Optional[str] = None
    metadata: Optional[dict[str, Any]] = None
    structured_content: Optional[list[dict]] = None
    image_url: Optional[str] = None  # For URL-based images
    image_data: Optional[dict[str, str]] = (
        None  # For base64 {media_type, data}
    )

    class Config:
        populate_by_name = True
        json_schema_extra = {
            "example": {
                "role": "user",
                "content": "This is a test message.",
                "name": None,
                "function_call": None,
                "tool_calls": None,
            }
        }