model-library 0.1.7__py3-none-any.whl → 0.1.9__py3-none-any.whl

model_library/providers/anthropic.py

@@ -1,16 +1,20 @@
+import datetime
 import io
 import logging
+import time
 from typing import Any, Literal, Sequence, cast
 
-from anthropic import AsyncAnthropic
+from anthropic import APIConnectionError, AsyncAnthropic
 from anthropic.types.beta.beta_tool_use_block import BetaToolUseBlock
 from anthropic.types.beta.parsed_beta_message import ParsedBetaMessage
+from pydantic import SecretStr
 from typing_extensions import override
 
 from model_library import model_library_settings
 from model_library.base import (
     LLM,
     BatchResult,
+    DelegateConfig,
     FileBase,
     FileInput,
     FileWithBase64,
@@ -19,9 +23,11 @@ from model_library.base import (
     InputItem,
     LLMBatchMixin,
     LLMConfig,
+    ProviderConfig,
     QueryResult,
     QueryResultCost,
     QueryResultMetadata,
+    RateLimit,
     RawInput,
     RawResponse,
     TextInput,
@@ -31,6 +37,7 @@ from model_library.base import (
     ToolResult,
 )
 from model_library.exceptions import (
+    ImmediateRetryException,
     MaxOutputTokensExceededError,
     NoMatchingToolCallError,
 )
@@ -38,11 +45,15 @@ from model_library.model_utils import get_default_budget_tokens
 from model_library.providers.openai import OpenAIModel
 from model_library.register_models import register_provider
 from model_library.utils import (
-    create_openai_client_with_defaults,
-    default_httpx_client,
+    create_anthropic_client_with_defaults,
 )
 
 
+class AnthropicConfig(ProviderConfig):
+    supports_compute_effort: bool = False
+    supports_auto_thinking: bool = False
+
+
 class AnthropicBatchMixin(LLMBatchMixin):
     """Batch processing support for Anthropic's Message Batches API."""
 
@@ -246,21 +257,27 @@ class AnthropicBatchMixin(LLMBatchMixin):
 
 @register_provider("anthropic")
 class AnthropicModel(LLM):
-    _client: AsyncAnthropic | None = None
+    provider_config = AnthropicConfig()
+
+    def _get_default_api_key(self) -> str:
+        if self.delegate_config:
+            return self.delegate_config.api_key.get_secret_value()
+        return model_library_settings.ANTHROPIC_API_KEY
 
     @override
-    def get_client(self) -> AsyncAnthropic:
-        if self._delegate_client:
-            return self._delegate_client
-        if not AnthropicModel._client:
+    def get_client(self, api_key: str | None = None) -> AsyncAnthropic:
+        if not self.has_client():
+            assert api_key
             headers: dict[str, str] = {}
-            AnthropicModel._client = AsyncAnthropic(
-                api_key=model_library_settings.ANTHROPIC_API_KEY,
-                http_client=default_httpx_client(),
-                max_retries=1,
+            client = create_anthropic_client_with_defaults(
+                base_url=self.delegate_config.base_url
+                if self.delegate_config
+                else None,
+                api_key=api_key,
                 default_headers=headers,
             )
-        return AnthropicModel._client
+            self.assign_client(client)
+        return super().get_client()
 
     def __init__(
         self,
@@ -268,33 +285,32 @@ class AnthropicModel(LLM):
         provider: str = "anthropic",
         *,
         config: LLMConfig | None = None,
-        custom_client: AsyncAnthropic | None = None,
+        delegate_config: DelegateConfig | None = None,
     ):
-        super().__init__(model_name, provider, config=config)
+        self.delegate_config = delegate_config
 
-        # allow custom client to act as delegate (native)
-        self._delegate_client: AsyncAnthropic | None = custom_client
+        super().__init__(model_name, provider, config=config)
 
         # https://docs.anthropic.com/en/api/openai-sdk
         self.delegate = (
             None
-            if self.native or custom_client
+            if self.native or self.delegate_config
             else OpenAIModel(
                 model_name=self.model_name,
-                provider=provider,
+                provider=self.provider,
                 config=config,
-                custom_client=create_openai_client_with_defaults(
-                    api_key=model_library_settings.ANTHROPIC_API_KEY,
+                use_completions=True,
+                delegate_config=DelegateConfig(
                     base_url="https://api.anthropic.com/v1/",
+                    api_key=SecretStr(model_library_settings.ANTHROPIC_API_KEY),
                 ),
-                use_completions=True,
             )
         )
 
         # Initialize batch support if enabled
         # Disable batch when using custom_client (similar to OpenAI)
         self.supports_batch: bool = (
-            self.supports_batch and self.native and not custom_client
+            self.supports_batch and self.native and not self.delegate_config
         )
         self.batch: LLMBatchMixin | None = (
             AnthropicBatchMixin(self) if self.supports_batch else None
@@ -520,7 +536,6 @@ class AnthropicModel(LLM):
         **kwargs: object,
     ) -> dict[str, Any]:
         body: dict[str, Any] = {
-            "max_tokens": self.max_tokens,
             "model": self.model_name,
             "messages": await self.parse_input(input),
         }
@@ -534,14 +549,28 @@ class AnthropicModel(LLM):
                 }
             ]
 
+        if not self.max_tokens:
+            raise Exception("Anthropic models require a max_tokens parameter")
+
+        body["max_tokens"] = self.max_tokens
+
         if self.reasoning:
-            budget_tokens = kwargs.pop(
-                "budget_tokens", get_default_budget_tokens(self.max_tokens)
-            )
-            body["thinking"] = {
-                "type": "enabled",
-                "budget_tokens": budget_tokens,
-            }
+            if self.provider_config.supports_auto_thinking:
+                body["thinking"] = {"type": "auto"}
+            else:
+                budget_tokens = kwargs.pop(
+                    "budget_tokens", get_default_budget_tokens(self.max_tokens)
+                )
+                body["thinking"] = {
+                    "type": "enabled",
+                    "budget_tokens": budget_tokens,
+                }
+
+        # effort controls compute allocation for text, tool calls, and thinking. Opus-4.5+
+        # use instead of reasoning_effort with auto_thinking
+        if self.provider_config.supports_compute_effort and self.compute_effort:
+            # default is "high"
+            body["output_config"] = {"effort": self.compute_effort}
 
         # Thinking models don't support temperature: https://docs.claude.com/en/docs/build-with-claude/extended-thinking#feature-compatibility
         if self.supports_temperature and not self.reasoning:
@@ -577,7 +606,7 @@ class AnthropicModel(LLM):
         client = self.get_client()
 
         # only send betas for the official Anthropic endpoint
-        is_anthropic_endpoint = self._delegate_client is None
+        is_anthropic_endpoint = self.delegate_config is None
         if not is_anthropic_endpoint:
             client_base_url = getattr(client, "_base_url", None) or getattr(
                 client, "base_url", None
@@ -587,16 +616,29 @@ class AnthropicModel(LLM):
 
         stream_kwargs = {**body}
         if is_anthropic_endpoint:
-            betas = ["files-api-2025-04-14", "interleaved-thinking-2025-05-14"]
+            betas = ["files-api-2025-04-14"]
+            if self.provider_config.supports_auto_thinking:
+                betas.extend(
+                    [
+                        "auto-thinking-2026-01-12",
+                        "effort-2025-11-24",
+                        "max-effort-2026-01-24",
+                    ]
+                )
+            else:
+                betas.extend(["interleaved-thinking-2025-05-14"])
             if "sonnet-4-5" in self.model_name:
                 betas.append("context-1m-2025-08-07")
             stream_kwargs["betas"] = betas
 
-        async with client.beta.messages.stream(
-            **stream_kwargs,
-        ) as stream:  # pyright: ignore[reportAny]
-            message = await stream.get_final_message()
-        self.logger.info(f"Anthropic Response finished: {message.id}")
+        try:
+            async with client.beta.messages.stream(
+                **stream_kwargs,
+            ) as stream:  # pyright: ignore[reportAny]
+                message = await stream.get_final_message()
+            self.logger.info(f"Anthropic Response finished: {message.id}")
+        except APIConnectionError:
+            raise ImmediateRetryException("Failed to connect to Anthropic")
 
         text = ""
         reasoning = ""
@@ -632,6 +674,38 @@ class AnthropicModel(LLM):
             history=[*input, RawResponse(response=message)],
         )
 
+    @override
+    async def get_rate_limit(self) -> RateLimit:
+        response = await self.get_client().messages.with_raw_response.create(
+            max_tokens=1,
+            messages=[
+                {
+                    "role": "user",
+                    "content": "Ping",
+                }
+            ],
+            model=self.model_name,
+        )
+        headers = response.headers
+
+        server_time_str = headers.get("date")
+        if server_time_str:
+            server_time = datetime.datetime.strptime(
+                server_time_str, "%a, %d %b %Y %H:%M:%S GMT"
+            ).replace(tzinfo=datetime.timezone.utc)
+            timestamp = server_time.timestamp()
+        else:
+            timestamp = time.time()
+
+        return RateLimit(
+            unix_timestamp=timestamp,
+            raw=headers,
+            request_limit=int(headers["anthropic-ratelimit-requests-limit"]),
+            request_remaining=int(headers["anthropic-ratelimit-requests-remaining"]),
+            token_limit=int(response.headers["anthropic-ratelimit-tokens-limit"]),
+            token_remaining=int(headers["anthropic-ratelimit-tokens-remaining"]),
+        )
+
     @override
     async def count_tokens(
         self,
@@ -645,20 +719,26 @@ class AnthropicModel(LLM):
         Count the number of tokens using Anthropic's native token counting API.
         https://docs.anthropic.com/en/docs/build-with-claude/token-counting
         """
-        input = [*history, *input]
-        if not input:
-            return 0
+        try:
+            input = [*history, *input]
+            if not input:
+                return 0
 
-        body = await self.build_body(input, tools=tools, **kwargs)
+            body = await self.build_body(input, tools=tools, **kwargs)
 
-        # Remove fields not supported by count_tokens endpoint
-        body.pop("max_tokens", None)
-        body.pop("temperature", None)
+            # Remove fields not supported by count_tokens endpoint
+            body.pop("max_tokens", None)
+            body.pop("temperature", None)
 
-        client = self.get_client()
-        response = await client.messages.count_tokens(**body)
+            client = self.get_client()
+            response = await client.messages.count_tokens(**body)
 
-        return response.input_tokens
+            return response.input_tokens
+        except Exception as e:
+            self.logger.error(f"Error counting tokens: {e}")
+            return await super().count_tokens(
+                input, history=history, tools=tools, **kwargs
+            )
 
     @override
     async def _calculate_cost(

model_library/providers/azure.py

@@ -1,3 +1,4 @@
+import json
 from typing import Literal
 
 from openai.lib.azure import AsyncAzureOpenAI
@@ -14,21 +15,32 @@ from model_library.utils import default_httpx_client
 
 @register_provider("azure")
 class AzureOpenAIModel(OpenAIModel):
-    _azure_client: AsyncAzureOpenAI | None = None
-
     @override
-    def get_client(self) -> AsyncAzureOpenAI:
-        if not AzureOpenAIModel._azure_client:
-            AzureOpenAIModel._azure_client = AsyncAzureOpenAI(
-                api_key=model_library_settings.AZURE_API_KEY,
-                azure_endpoint=model_library_settings.AZURE_ENDPOINT,
-                api_version=model_library_settings.get(
+    def _get_default_api_key(self) -> str:
+        return json.dumps(
+            {
+                "AZURE_API_KEY": model_library_settings.AZURE_API_KEY,
+                "AZURE_ENDPOINT": model_library_settings.AZURE_ENDPOINT,
+                "AZURE_API_VERSION": model_library_settings.get(
                     "AZURE_API_VERSION", "2025-04-01-preview"
                 ),
+            }
+        )
+
+    @override
+    def get_client(self, api_key: str | None = None) -> AsyncAzureOpenAI:
+        if not self.has_client():
+            assert api_key
+            creds = json.loads(api_key)
+            client = AsyncAzureOpenAI(
+                api_key=creds["AZURE_API_KEY"],
+                azure_endpoint=creds["AZURE_ENDPOINT"],
+                api_version=creds["AZURE_API_VERSION"],
                 http_client=default_httpx_client(),
-                max_retries=1,
+                max_retries=3,
             )
-        return AzureOpenAIModel._azure_client
+            self.assign_client(client)
+        return super(OpenAIModel, self).get_client(api_key)
 
     def __init__(
         self,

model_library/providers/cohere.py

@@ -1,13 +1,14 @@
 from typing import Literal
 
+from pydantic import SecretStr
+
 from model_library import model_library_settings
 from model_library.base import (
+    DelegateConfig,
     DelegateOnly,
     LLMConfig,
 )
-from model_library.providers.openai import OpenAIModel
 from model_library.register_models import register_provider
-from model_library.utils import create_openai_client_with_defaults
 
 
 @register_provider("cohere")
@@ -22,13 +23,12 @@ class CohereModel(DelegateOnly):
         super().__init__(model_name, provider, config=config)
 
         # https://docs.cohere.com/docs/compatibility-api
-        self.delegate = OpenAIModel(
-            model_name=self.model_name,
-            provider=self.provider,
+        self.init_delegate(
             config=config,
-            custom_client=create_openai_client_with_defaults(
-                api_key=model_library_settings.COHERE_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://api.cohere.ai/compatibility/v1",
+                api_key=SecretStr(model_library_settings.COHERE_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )

model_library/providers/deepseek.py

@@ -5,14 +5,15 @@ https://cdn.deepseek.com/policies/en-US/deepseek-privacy-policy.html
 
 from typing import Literal
 
+from pydantic import SecretStr
+
 from model_library import model_library_settings
 from model_library.base import (
+    DelegateConfig,
     DelegateOnly,
     LLMConfig,
 )
-from model_library.providers.openai import OpenAIModel
 from model_library.register_models import register_provider
-from model_library.utils import create_openai_client_with_defaults
 
 
 @register_provider("deepseek")
@@ -27,13 +28,12 @@ class DeepSeekModel(DelegateOnly):
         super().__init__(model_name, provider, config=config)
 
         # https://api-docs.deepseek.com/
-        self.delegate = OpenAIModel(
-            model_name=self.model_name,
-            provider=self.provider,
+        self.init_delegate(
             config=config,
-            custom_client=create_openai_client_with_defaults(
-                api_key=model_library_settings.DEEPSEEK_API_KEY,
-                base_url="https://api.deepseek.com",
+            delegate_config=DelegateConfig(
+                base_url="https://api.deepseek.com/v1",
+                api_key=SecretStr(model_library_settings.DEEPSEEK_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )

model_library/providers/fireworks.py

@@ -1,18 +1,18 @@
 from typing import Literal
 
+from pydantic import SecretStr
 from typing_extensions import override
 
 from model_library import model_library_settings
 from model_library.base import (
+    DelegateConfig,
+    DelegateOnly,
     LLMConfig,
     ProviderConfig,
     QueryResultCost,
     QueryResultMetadata,
 )
-from model_library.base.delegate_only import DelegateOnly
-from model_library.providers.openai import OpenAIModel
 from model_library.register_models import register_provider
-from model_library.utils import create_openai_client_with_defaults
 
 
 class FireworksConfig(ProviderConfig):
@@ -38,15 +38,14 @@ class FireworksModel(DelegateOnly):
             self.model_name = "accounts/rayan-936e28/deployedModels/" + self.model_name
 
         # https://docs.fireworks.ai/tools-sdks/openai-compatibility
-        self.delegate = OpenAIModel(
-            model_name=self.model_name,
-            provider=self.provider,
+        self.init_delegate(
             config=config,
-            custom_client=create_openai_client_with_defaults(
-                api_key=model_library_settings.FIREWORKS_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://api.fireworks.ai/inference/v1",
+                api_key=SecretStr(model_library_settings.FIREWORKS_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )
 
     @override

model_library/providers/google/batch.py

@@ -24,16 +24,19 @@ from google.genai.types import (
 )
 
 
-def extract_text_from_json_response(response: dict[str, Any]) -> str:
+def extract_text_from_json_response(response: dict[str, Any]) -> tuple[str, str]:
     """Extract concatenated non-thought text from a JSON response structure."""
     # TODO: fix the typing we always ignore
     text = ""
+    reasoning = ""
     for candidate in response.get("candidates", []) or []:  # type: ignore
         content = (candidate or {}).get("content") or {}  # type: ignore
         for part in content.get("parts", []) or []:  # type: ignore
-            if not part.get("thought", False):  # type: ignore
+            if part.get("thought", False):  # type: ignore
+                reasoning += part.get("text", "")  # type: ignore
+            else:
                 text += part.get("text", "")  # type: ignore
-    return text  # type: ignore
+    return text, reasoning  # type: ignore
 
 
 def parse_predictions_jsonl(jsonl: str) -> list[BatchResult]:
@@ -48,9 +51,10 @@ def parse_predictions_jsonl(jsonl: str) -> list[BatchResult]:
         custom_id = data.get("key", "unknown")
         if "response" in data:
             response = data["response"]
-            text = extract_text_from_json_response(response)
+            text, reasoning = extract_text_from_json_response(response)
             output = QueryResult()
             output.output_text = text
+            output.reasoning = reasoning
             if "usageMetadata" in response:
                 output.metadata.in_tokens = response["usageMetadata"].get(
                     "promptTokenCount", 0
@@ -196,7 +200,7 @@ class GoogleBatchMixin(LLMBatchMixin):
             custom_id = labels.get("qa_pair_id", f"request-{i}")
             jsonl_lines.append(json.dumps({"key": custom_id, "request": request_data}))
 
-        batch_request_file = self._root.client.files.upload(
+        batch_request_file = self._root.get_client().files.upload(
             file=io.StringIO("\n".join(jsonl_lines)),
             config=UploadFileConfig(mime_type="application/jsonl"),
         )
@@ -205,7 +209,7 @@ class GoogleBatchMixin(LLMBatchMixin):
             raise Exception("Failed to upload batch jsonl")
 
         try:
-            job: BatchJob = await self._root.client.aio.batches.create(
+            job: BatchJob = await self._root.get_client().aio.batches.create(
                 model=self._root.model_name,
                 src=batch_request_file.name,
                 config={"display_name": batch_name},
@@ -224,14 +228,14 @@ class GoogleBatchMixin(LLMBatchMixin):
     async def get_batch_results(self, batch_id: str) -> list[BatchResult]:
         self._root.logger.info(f"Retrieving batch results for {batch_id}")
 
-        job = await self._root.client.aio.batches.get(name=batch_id)
+        job = await self._root.get_client().aio.batches.get(name=batch_id)
 
         results: list[BatchResult] = []
 
         if job.state == JobState.JOB_STATE_SUCCEEDED:
             if job.dest and job.dest.file_name:
                 results_file_name = job.dest.file_name
-                file_content = await self._root.client.aio.files.download(
+                file_content = await self._root.get_client().aio.files.download(
                     file=results_file_name
                 )
                 decoded = file_content.decode("utf-8")
@@ -250,7 +254,7 @@ class GoogleBatchMixin(LLMBatchMixin):
     @override
     async def cancel_batch_request(self, batch_id: str):
         self._root.logger.info(f"Cancelling batch {batch_id}")
-        await self._root.client.aio.batches.cancel(name=batch_id)
+        await self._root.get_client().aio.batches.cancel(name=batch_id)
 
     @override
     async def get_batch_progress(self, batch_id: str) -> int:
@@ -262,7 +266,7 @@ class GoogleBatchMixin(LLMBatchMixin):
 
         try:
             self._root.logger.debug(f"Checking batch status for {batch_id}")
-            job: BatchJob = await self._root.client.aio.batches.get(name=batch_id)
+            job: BatchJob = await self._root.get_client().aio.batches.get(name=batch_id)
             state = job.state
 
             if not state: