Implements LLMProvider + BatchProvider for Gemini.
| 23 | |
| 24 | |
| 25 | class GeminiProvider: |
| 26 | """Implements LLMProvider + BatchProvider for Gemini.""" |
| 27 | |
| 28 | def __init__(self, model: str, *, reasoning_effort: str | None = None) -> None: |
| 29 | self._model = model |
| 30 | self._gemini_model = model.removeprefix("gemini/") |
| 31 | self._thinking_budget: int | None = ( |
| 32 | int(reasoning_effort) if reasoning_effort is not None else None |
| 33 | ) |
| 34 | self.client = genai.Client(api_key=os.environ.get("GEMINI_API_KEY")) |
| 35 | |
| 36 | def _thinking_config(self) -> types.ThinkingConfig | None: |
| 37 | if self._thinking_budget is None: |
| 38 | return None |
| 39 | return types.ThinkingConfig(thinking_budget=self._thinking_budget) |
| 40 | |
| 41 | def call(self, user_content: str) -> tuple[str, TokenUsage]: |
| 42 | config_kwargs: dict = { |
| 43 | "system_instruction": SYSTEM_PROMPT, |
| 44 | "response_mime_type": "application/json", |
| 45 | "http_options": types.HttpOptions(timeout=LLM_TIMEOUT_SECONDS * 1000), |
| 46 | } |
| 47 | thinking = self._thinking_config() |
| 48 | if thinking: |
| 49 | config_kwargs["thinking_config"] = thinking |
| 50 | response = self.client.models.generate_content( |
| 51 | model=self._gemini_model, |
| 52 | contents=user_content, |
| 53 | config=types.GenerateContentConfig(**config_kwargs), |
| 54 | ) |
| 55 | usage = TokenUsage() |
| 56 | um = response.usage_metadata |
| 57 | if um: |
| 58 | usage.input_tokens = um.prompt_token_count or 0 |
| 59 | usage.output_tokens = um.candidates_token_count or 0 |
| 60 | usage.reasoning_tokens = um.thoughts_token_count or 0 |
| 61 | return response.text, usage |
| 62 | |
| 63 | @property |
| 64 | def retryable_exceptions(self) -> tuple[type[Exception], ...]: |
| 65 | return (ClientError, ServerError, httpx.TimeoutException) |
| 66 | |
| 67 | # -- Batch API -- |
| 68 | |
| 69 | def submit_batch(self, entries: list[BatchEntry]) -> str: |
| 70 | config_kwargs: dict = { |
| 71 | "system_instruction": SYSTEM_PROMPT, |
| 72 | "response_mime_type": "application/json", |
| 73 | } |
| 74 | thinking = self._thinking_config() |
| 75 | if thinking: |
| 76 | config_kwargs["thinking_config"] = thinking |
| 77 | |
| 78 | inline_entries = [] |
| 79 | for entry in entries: |
| 80 | inline_entries.append( |
| 81 | types.InlinedRequest( |
| 82 | contents=entry.user_content, |
no outgoing calls
no test coverage detected