base_llm.py

# Copyright 2025 Google LLC
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

from __future__ import annotations

from abc import abstractmethod
from typing import AsyncGenerator
from typing import TYPE_CHECKING

from google.genai import types
from pydantic import BaseModel
from pydantic import ConfigDict

from .base_llm_connection import BaseLlmConnection

if TYPE_CHECKING:
  from .llm_request import LlmRequest
  from .llm_response import LlmResponse


class BaseLlm(BaseModel):
  """The BaseLLM class."""

  model_config = ConfigDict(
      # This allows us to use arbitrary types in the model. E.g. PIL.Image.
      arbitrary_types_allowed=True,
  )
  """The pydantic model config."""

  model: str
  """The name of the LLM, e.g. gemini-2.5-flash or gemini-2.5-pro."""

  @classmethod
  def supported_models(cls) -> list[str]:
    """Returns a list of supported models in regex for LlmRegistry."""
    return []

  @abstractmethod
  async def generate_content_async(
      self, llm_request: LlmRequest, stream: bool = False
  ) -> AsyncGenerator[LlmResponse, None]:
    """Generates content for a single model turn.

    This method handles Server-Sent Events (SSE) streaming for unidirectional
    content generation. For bidirectional streaming (e.g., Gemini Live API),
    use the `connect()` method instead.

    Args:
      llm_request: LlmRequest, the request to send to the LLM.
      stream: bool = False, whether to enable SSE streaming mode.

    Yields:
      LlmResponse objects representing the model's response for one turn.

      **Non-streaming mode (stream=False):**

        Yields exactly one LlmResponse containing the complete model output
        (text, function calls, bytes, etc.). This response has `partial=False`.

      **Streaming mode (stream=True):**

        Yields multiple LlmResponse objects as chunks arrive:

        - Intermediate chunks: `partial=True` (progressive updates)
        - Final chunk: `partial=False` (aggregated content from entire turn,
          identical to stream=False output)
        - Text consolidation: Consecutive text parts of the same type
          (thought/non-thought) SHOULD merge without separator, but client
          code must not rely on this - unconsolidated parts are unusual but also
          valid

      **Common content in partial chunks:**

        All intermediate chunks have `partial=True` regardless of content type.
        Common examples include:

        - Text: Streams incrementally as tokens arrive
        - Function calls: May arrive in separate chunks
        - Bytes (e.g., images): Typically arrive as single chunk, interleaved
          with text
        - Thoughts: Stream incrementally when thinking_config is enabled

      **Examples:**

      1. Simple text streaming::

           LlmResponse(partial=True,  parts=["The weather"])
           LlmResponse(partial=True,  parts=[" in Tokyo is"])
           LlmResponse(partial=True,  parts=[" sunny."])
           LlmResponse(partial=False, parts=["The weather in Tokyo is sunny."])

      2. Text + function call::

           LlmResponse(partial=True,  parts=[Text("Let me check...")])
           LlmResponse(partial=True,  parts=[FunctionCall("get_weather", ...)])
           LlmResponse(partial=False, parts=[Text("Let me check..."),
                                             FunctionCall("get_weather", ...)])

      3. Parallel function calls across chunks::

           LlmResponse(partial=True,  parts=[Text("Checking both cities...")])
           LlmResponse(partial=True,  parts=[FunctionCall("get_weather", Tokyo)])
           LlmResponse(partial=True,  parts=[FunctionCall("get_weather", NYC)])
           LlmResponse(partial=False, parts=[Text("Checking both cities..."),
                                             FunctionCall("get_weather", Tokyo),
                                             FunctionCall("get_weather", NYC)])

      4. Text + bytes (image generation with gemini-2.5-flash-image)::

           LlmResponse(partial=True,  parts=[Text("Here's an image of a dog.")])
           LlmResponse(partial=True,  parts=[Text("\n")])
           LlmResponse(partial=True,  parts=[Blob(image/png, 1.6MB)])
           LlmResponse(partial=True,  parts=[Text("It carries a bone")])
           LlmResponse(partial=True,  parts=[Text(" and running around.")])
           LlmResponse(partial=False, parts=[Text("Here's an image of a dog.\n"),
                                             Blob(image/png, 1.6MB),
                                             Text("It carries a bone and running around.")])

         Note: Consecutive text parts before and after blob merge separately.

      5. Text with thinking (gemini-2.5-flash with thinking_config)::

           LlmResponse(partial=True,  parts=[Thought("Let me analyze...")])
           LlmResponse(partial=True,  parts=[Thought("The user wants...")])
           LlmResponse(partial=True,  parts=[Text("Based on my analysis,")])
           LlmResponse(partial=True,  parts=[Text(" the answer is 42.")])
           LlmResponse(partial=False, parts=[Thought("Let me analyze...The user wants..."),
                                             Text("Based on my analysis, the answer is 42.")])

         Note: Consecutive parts of same type merge (thoughts→thought, text→text).

      **Important:** All yielded responses represent one logical model turn.
      The final response with `partial=False` should be identical to the
      response that would be received with `stream=False`.
    """
    raise NotImplementedError(
        f'Async generation is not supported for {self.model}.'
    )
    yield  # AsyncGenerator requires a yield statement in function body.

  def _maybe_append_user_content(self, llm_request: LlmRequest):
    """Appends a user content, so that model can continue to output.

    Args:
      llm_request: LlmRequest, the request to send to the Gemini model.
    """
    # If no content is provided, append a user content to hint model response
    # using system instruction.
    if not llm_request.contents:
      llm_request.contents.append(
          types.Content(
              role='user',
              parts=[
                  types.Part(
                      text=(
                          'Handle the requests as specified in the System'
                          ' Instruction.'
                      )
                  )
              ],
          )
      )
      return

    # Insert a user content to preserve user intent and to avoid empty
    # model response.
    if llm_request.contents[-1].role != 'user':
      llm_request.contents.append(
          types.Content(
              role='user',
              parts=[
                  types.Part(
                      text=(
                          'Continue processing previous requests as instructed.'
                          ' Exit or provide a summary if no more outputs are'
                          ' needed.'
                      )
                  )
              ],
          )
      )

  def connect(self, llm_request: LlmRequest) -> BaseLlmConnection:
    """Creates a live connection to the LLM.

    Args:
      llm_request: LlmRequest, the request to send to the LLM.

    Returns:
      BaseLlmConnection, the connection to the LLM.
    """
    raise NotImplementedError(
        f'Live connection is not supported for {self.model}.'
    )

  async def aclose(self) -> None:
    """Closes the LLM and releases resources.

    This method provides a lifecycle hook for cleanup when the LLM is no longer
    needed. The default implementation is a no-op for backward compatibility.

    Subclasses that manage resources (e.g., HTTP clients) should override this
    method to perform proper cleanup.
    """
    pass