model-library 0.1.7__py3-none-any.whl → 0.1.8__py3-none-any.whl

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:

model_library/providers/google/google.py

@@ -1,5 +1,6 @@
 import base64
 import io
+import json
 import logging
 import uuid
 from typing import Any, Literal, Sequence, cast
@@ -25,6 +26,7 @@ from google.genai.types import (
     ToolListUnion,
     UploadFileConfig,
 )
+from google.oauth2 import service_account
 from typing_extensions import override
 
 from model_library import model_library_settings
@@ -95,31 +97,50 @@ class GoogleModel(LLM):
         ),
     ]
 
-    @override
-    def get_client(self) -> Client:
-        if self.provider_config.use_vertex:
-            # Preview Gemini releases from September 2025 are only served from the global
-            # Vertex region. The public docs for these SKUs list `global` as the sole
-            # availability region (see September 25, 2025 release notes), so we override
-            # the default `us-central1` when we detect them.
-            # https://cloud.google.com/vertex-ai/generative-ai/docs/models/gemini/2-5-flash
-            MODEL_REGION_OVERRIDES: dict[str, str] = {
-                "gemini-2.5-flash-preview-09-2025": "global",
-                "gemini-2.5-flash-lite-preview-09-2025": "global",
+    def _get_default_api_key(self) -> str:
+        if not self.provider_config.use_vertex:
+            return model_library_settings.GOOGLE_API_KEY
+
+        return json.dumps(
+            {
+                "GCP_REGION": model_library_settings.GCP_REGION,
+                "GCP_PROJECT_ID": model_library_settings.GCP_PROJECT_ID,
+                "GCP_CREDS": model_library_settings.GCP_CREDS,
             }
-            region = model_library_settings.GCP_REGION
-            if self.model_name in MODEL_REGION_OVERRIDES:
-                region = MODEL_REGION_OVERRIDES[self.model_name]
-
-            return Client(
-                vertexai=True,
-                project=model_library_settings.GCP_PROJECT_ID,
-                location=region,
-                # Credentials object is not typed, so we have to ignore the error
-                credentials=model_library_settings.GCP_CREDS,
-            )
+        )
 
-        return Client(api_key=model_library_settings.GOOGLE_API_KEY)
+    @override
+    def get_client(self, api_key: str | None = None) -> Client:
+        if not self.has_client():
+            assert api_key
+            if self.provider_config.use_vertex:
+                # Gemini preview releases are only server from the global Vertex region after September 2025.
+                MODEL_REGION_OVERRIDES: dict[str, str] = {
+                    "gemini-2.5-flash-preview-09-2025": "global",
+                    "gemini-2.5-flash-lite-preview-09-2025": "global",
+                    "gemini-3-flash-preview": "global",
+                    "gemini-3-pro-preview": "global",
+                }
+
+                creds = json.loads(api_key)
+
+                region = creds["GCP_REGION"]
+                if self.model_name in MODEL_REGION_OVERRIDES:
+                    region = MODEL_REGION_OVERRIDES[self.model_name]
+
+                client = Client(
+                    vertexai=True,
+                    project=creds["GCP_PROJECT_ID"],
+                    location=region,
+                    credentials=service_account.Credentials.from_service_account_info(  # type: ignore
+                        json.loads(creds["GCP_CREDS"]),
+                        scopes=["https://www.googleapis.com/auth/cloud-platform"],
+                    ),
+                )
+            else:
+                client = Client(api_key=api_key)
+            self.assign_client(client)
+        return super().get_client()
 
     def __init__(
         self,
@@ -141,8 +162,6 @@ class GoogleModel(LLM):
             GoogleBatchMixin(self) if self.supports_batch else None
         )
 
-        self.client = self.get_client()
-
     @override
     async def parse_input(
         self,
@@ -260,7 +279,7 @@ class GoogleModel(LLM):
             )
 
         mime = f"image/{mime}" if type == "image" else mime  # TODO:
-        response: File = self.client.files.upload(
+        response: File = self.get_client().files.upload(
             file=bytes, config=UploadFileConfig(mime_type=mime)
         )
         if not response.name:
@@ -338,7 +357,7 @@ class GoogleModel(LLM):
 
         metadata: GenerateContentResponseUsageMetadata | None = None
 
-        stream = await self.client.aio.models.generate_content_stream(**body)
+        stream = await self.get_client().aio.models.generate_content_stream(**body)
         contents: list[Content | None] = []
         finish_reason: FinishReason | None = None
 
@@ -437,7 +456,7 @@ class GoogleModel(LLM):
             tools=parsed_tools,
         )
 
-        response = await self.client.aio.models.count_tokens(
+        response = await self.get_client().aio.models.count_tokens(
             model=self.model_name,
             contents=cast(Any, contents),
             config=config,
@@ -503,7 +522,7 @@ class GoogleModel(LLM):
         # Make the request with retry wrapper
         async def _query():
             try:
-                return await self.client.aio.models.generate_content(**body)
+                return await self.get_client().aio.models.generate_content(**body)
             except (genai_errors.ServerError, genai_errors.UnknownApiResponseError):
                 raise ImmediateRetryException("Failed to connect to Google API")
 

model_library/providers/inception.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("inception")
@@ -22,13 +23,12 @@ class MercuryModel(DelegateOnly):
         super().__init__(model_name, provider, config=config)
 
         # https://docs.inceptionlabs.ai/get-started/get-started#external-libraries-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.MERCURY_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://api.inceptionlabs.ai/v1/",
+                api_key=SecretStr(model_library_settings.MERCURY_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )

model_library/providers/kimi.py

@@ -1,13 +1,16 @@
-from typing import Literal
+from typing import Any, Literal
+
+from typing_extensions import override
+
+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("kimi")
@@ -22,13 +25,20 @@ class KimiModel(DelegateOnly):
         super().__init__(model_name, provider, config=config)
 
         # https://platform.moonshot.ai/docs/guide/migrating-from-openai-to-kimi#about-api-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.KIMI_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://api.moonshot.ai/v1/",
+                api_key=SecretStr(model_library_settings.KIMI_API_KEY),
             ),
             use_completions=True,
+            delegate_provider="openai",
         )
+
+    @override
+    def _get_extra_body(self) -> dict[str, Any]:
+        """
+        Build extra body parameters for Kimi-specific features.
+        see https://platform.moonshot.ai/docs/guide/kimi-k2-5-quickstart#parameters-differences-in-request-body
+        """
+        return {"thinking": {"type": "enabled" if self.reasoning else "disabled"}}

model_library/providers/minimax.py

@@ -1,16 +1,17 @@
-from typing import Literal
+from typing import Literal, Sequence
+
+from pydantic import SecretStr
+from typing_extensions import override
 
 from model_library import model_library_settings
-from model_library.base import DelegateOnly, LLMConfig
-from model_library.base.input import InputItem, ToolDefinition
-from model_library.providers.anthropic import AnthropicModel
+from model_library.base import (
+    DelegateConfig,
+    DelegateOnly,
+    InputItem,
+    LLMConfig,
+    ToolDefinition,
+)
 from model_library.register_models import register_provider
-from model_library.utils import default_httpx_client
-
-from anthropic import AsyncAnthropic
-
-from typing import Sequence
-from typing_extensions import override
 
 
 @register_provider("minimax")
@@ -24,16 +25,13 @@ class MinimaxModel(DelegateOnly):
     ):
         super().__init__(model_name, provider, config=config)
 
-        self.delegate = AnthropicModel(
-            model_name=self.model_name,
-            provider=self.provider,
+        self.init_delegate(
             config=config,
-            custom_client=AsyncAnthropic(
-                api_key=model_library_settings.MINIMAX_API_KEY,
+            delegate_config=DelegateConfig(
                 base_url="https://api.minimax.io/anthropic",
-                http_client=default_httpx_client(),
-                max_retries=1,
+                api_key=SecretStr(model_library_settings.MINIMAX_API_KEY),
             ),
+            delegate_provider="anthropic",
         )
 
     # minimax client shares anthropic's syntax

model_library/providers/mistral.py

@@ -3,7 +3,13 @@ import logging
 from collections.abc import Sequence
 from typing import Any, Literal
 
-from mistralai import AssistantMessage, ContentChunk, Mistral, TextChunk, ThinkChunk
+from mistralai import (
+    AssistantMessage,
+    ContentChunk,
+    Mistral,
+    TextChunk,
+    ThinkChunk,
+)
 from mistralai.models.completionevent import CompletionEvent
 from mistralai.models.toolcall import ToolCall as MistralToolCall
 from mistralai.utils.eventstreaming import EventStreamAsync
@@ -40,16 +46,20 @@ from model_library.utils import default_httpx_client
 
 @register_provider("mistralai")
 class MistralModel(LLM):
-    _client: Mistral | None = None
+    @override
+    def _get_default_api_key(self) -> str:
+        return model_library_settings.MISTRAL_API_KEY
 
     @override
-    def get_client(self) -> Mistral:
-        if not MistralModel._client:
-            MistralModel._client = Mistral(
-                api_key=model_library_settings.MISTRAL_API_KEY,
+    def get_client(self, api_key: str | None = None) -> Mistral:
+        if not self.has_client():
+            assert api_key
+            client = Mistral(
+                api_key=api_key,
                 async_client=default_httpx_client(),
             )
-        return MistralModel._client
+            self.assign_client(client)
+        return super().get_client()
 
     def __init__(
         self,
@@ -198,12 +208,14 @@ class MistralModel(LLM):
 
         body: dict[str, Any] = {
             "model": self.model_name,
-            "max_tokens": self.max_tokens,
             "messages": messages,
             "prompt_mode": "reasoning" if self.reasoning else None,
             "tools": tools,
         }
 
+        if self.max_tokens:
+            body["max_tokens"] = self.max_tokens
+
         if self.supports_temperature:
             if self.temperature is not None:
                 body["temperature"] = self.temperature