model-library 0.1.5__py3-none-any.whl → 0.1.7__py3-none-any.whl

model_library/providers/amazon.py

@@ -3,6 +3,7 @@ import asyncio
 import base64
 import io
 import json
+import logging
 from typing import Any, Literal, Sequence, cast
 
 import boto3
@@ -12,24 +13,26 @@ from typing_extensions import override
 
 from model_library.base import (
     LLM,
+    FileBase,
     FileInput,
     FileWithBase64,
     FileWithId,
-    FileWithUrl,
     InputItem,
     LLMConfig,
     QueryResult,
     QueryResultMetadata,
+    RawInput,
+    RawResponse,
     TextInput,
     ToolBody,
     ToolCall,
     ToolDefinition,
     ToolResult,
 )
-from model_library.base.input import FileBase
 from model_library.exceptions import (
     BadInputError,
     MaxOutputTokensExceededError,
+    NoMatchingToolCallError,
 )
 from model_library.model_utils import get_default_budget_tokens
 from model_library.register_models import register_provider
@@ -69,6 +72,20 @@ class AmazonModel(LLM):
 
     cache_control = {"type": "default"}
 
+    async def get_tool_call_ids(self, input: Sequence[InputItem]) -> list[str]:
+        raw_responses = [x for x in input if isinstance(x, RawResponse)]
+        tool_call_ids: list[str] = []
+
+        calls = [
+            y["toolUse"]
+            for x in raw_responses
+            if "content" in x.response
+            for y in x.response["content"]
+            if "toolUse" in y
+        ]
+        tool_call_ids.extend([x["toolUseId"] for x in calls])
+        return tool_call_ids
+
     @override
     async def parse_input(
         self,
@@ -76,58 +93,63 @@ class AmazonModel(LLM):
         **kwargs: Any,
     ) -> list[dict[str, Any]]:
         new_input: list[dict[str, Any] | Any] = []
+
         content_user: list[dict[str, Any]] = []
 
+        def flush_content_user():
+            if content_user:
+                # NOTE: must make new object as we clear()
+                new_input.append({"role": "user", "content": content_user.copy()})
+                content_user.clear()
+
+        tool_call_ids = await self.get_tool_call_ids(input)
+
         for item in input:
+            if isinstance(item, TextInput):
+                content_user.append({"text": item.text})
+                continue
+
+            if isinstance(item, FileBase):
+                match item.type:
+                    case "image":
+                        parsed = await self.parse_image(item)
+                    case "file":
+                        parsed = await self.parse_file(item)
+                content_user.append(parsed)
+                continue
+
+            # non content user item
+            flush_content_user()
+
             match item:
-                case TextInput():
-                    content_user.append({"text": item.text})
-                case FileWithBase64() | FileWithUrl() | FileWithId():
-                    match item.type:
-                        case "image":
-                            content_user.append(await self.parse_image(item))
-                        case "file":
-                            content_user.append(await self.parse_file(item))
-                case _:
-                    if content_user:
-                        new_input.append({"role": "user", "content": content_user})
-                        content_user = []
-                    match item:
-                        case ToolResult():
-                            if not (
-                                isinstance(x, dict)
-                                and "toolUse" in x
-                                and x["toolUse"].get("toolUseId")
-                                == item.tool_call.call_id
-                                for x in new_input
-                            ):
-                                raise Exception(
-                                    "Tool call result provided with no matching tool call"
-                                )
-                            new_input.append(
+                case ToolResult():
+                    if item.tool_call.id not in tool_call_ids:
+                        raise NoMatchingToolCallError()
+
+                    new_input.append(
+                        {
+                            "role": "user",
+                            "content": [
                                 {
-                                    "role": "user",
-                                    "content": [
-                                        {
-                                            "toolResult": {
-                                                "toolUseId": item.tool_call.id,
-                                                "content": [
-                                                    {"json": {"result": item.result}}
-                                                ],
-                                            }
-                                        }
-                                    ],
+                                    "toolResult": {
+                                        "toolUseId": item.tool_call.id,
+                                        "content": [{"json": {"result": item.result}}],
+                                    }
                                 }
-                            )
-                        case dict():  # RawInputItem and RawResponse
-                            new_input.append(item)
+                            ],
+                        }
+                    )
+                case RawResponse():
+                    new_input.append(item.response)
+                case RawInput():
+                    new_input.append(item.input)
 
-        if content_user:
-            if self.supports_cache:
-                if not isinstance(input[-1], FileBase):
-                    # last item cannot be file
-                    content_user.append({"cachePoint": self.cache_control})
-            new_input.append({"role": "user", "content": content_user})
+        if content_user and self.supports_cache:
+            if not isinstance(input[-1], FileBase):
+                # last item cannot be file
+                content_user.append({"cachePoint": self.cache_control})
+
+        flush_content_user()
 
         return new_input
 
@@ -195,6 +217,7 @@ class AmazonModel(LLM):
     ) -> FileWithId:
         raise NotImplementedError()
 
+    @override
     async def build_body(
         self,
         input: Sequence[InputItem],
@@ -337,6 +360,7 @@ class AmazonModel(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         body = await self.build_body(input, tools=tools, **kwargs)
@@ -381,5 +405,5 @@ class AmazonModel(LLM):
             reasoning=reasoning,
             metadata=metadata,
             tool_calls=tool_calls,
-            history=[*input, messages],
+            history=[*input, RawResponse(response=messages)],
         )

model_library/providers/anthropic.py

@@ -1,16 +1,17 @@
 import io
+import logging
 from typing import Any, Literal, Sequence, cast
 
 from anthropic import AsyncAnthropic
-from anthropic.types import TextBlock, ToolUseBlock
 from anthropic.types.beta.beta_tool_use_block import BetaToolUseBlock
-from anthropic.types.message import Message
+from anthropic.types.beta.parsed_beta_message import ParsedBetaMessage
 from typing_extensions import override
 
 from model_library import model_library_settings
 from model_library.base import (
     LLM,
     BatchResult,
+    FileBase,
     FileInput,
     FileWithBase64,
     FileWithId,
@@ -21,7 +22,8 @@ from model_library.base import (
     QueryResult,
     QueryResultCost,
     QueryResultMetadata,
-    RawInputItem,
+    RawInput,
+    RawResponse,
     TextInput,
     ToolBody,
     ToolCall,
@@ -30,6 +32,7 @@ from model_library.base import (
 )
 from model_library.exceptions import (
     MaxOutputTokensExceededError,
+    NoMatchingToolCallError,
 )
 from model_library.model_utils import get_default_budget_tokens
 from model_library.providers.openai import OpenAIModel
@@ -37,8 +40,6 @@ from model_library.register_models import register_provider
 from model_library.utils import (
     create_openai_client_with_defaults,
     default_httpx_client,
-    filter_empty_text_blocks,
-    normalize_tool_result,
 )
 
 
@@ -61,9 +62,9 @@ class AnthropicBatchMixin(LLMBatchMixin):
 
         Format: {"custom_id": str, "params": {...message params...}}
         """
-        # Build the message body using the parent model's create_body method
+        # Build the message body using the parent model's build_body method
         tools = cast(list[ToolDefinition], kwargs.pop("tools", []))
-        body = await self._root.create_body(input, tools=tools, **kwargs)
+        body = await self._root.build_body(input, tools=tools, **kwargs)
 
         return {
             "custom_id": custom_id,
@@ -249,6 +250,8 @@ class AnthropicModel(LLM):
 
     @override
     def get_client(self) -> AsyncAnthropic:
+        if self._delegate_client:
+            return self._delegate_client
         if not AnthropicModel._client:
             headers: dict[str, str] = {}
             AnthropicModel._client = AsyncAnthropic(
@@ -262,16 +265,20 @@ class AnthropicModel(LLM):
     def __init__(
         self,
         model_name: str,
-        provider: Literal["anthropic"] = "anthropic",
+        provider: str = "anthropic",
         *,
         config: LLMConfig | None = None,
+        custom_client: AsyncAnthropic | None = None,
     ):
         super().__init__(model_name, provider, config=config)
 
+        # allow custom client to act as delegate (native)
+        self._delegate_client: AsyncAnthropic | None = custom_client
+
         # https://docs.anthropic.com/en/api/openai-sdk
-        self.delegate: OpenAIModel | None = (
+        self.delegate = (
             None
-            if self.native
+            if self.native or custom_client
             else OpenAIModel(
                 model_name=self.model_name,
                 provider=provider,
@@ -285,11 +292,28 @@ class AnthropicModel(LLM):
         )
 
         # Initialize batch support if enabled
-        self.supports_batch: bool = self.supports_batch and self.native
+        # Disable batch when using custom_client (similar to OpenAI)
+        self.supports_batch: bool = (
+            self.supports_batch and self.native and not custom_client
+        )
         self.batch: LLMBatchMixin | None = (
             AnthropicBatchMixin(self) if self.supports_batch else None
         )
 
+    async def get_tool_call_ids(self, input: Sequence[InputItem]) -> list[str]:
+        raw_responses = [x for x in input if isinstance(x, RawResponse)]
+        tool_call_ids: list[str] = []
+
+        calls = [
+            y
+            for x in raw_responses
+            if isinstance(x.response, ParsedBetaMessage)
+            for y in x.response.content  # pyright: ignore[reportUnknownMemberType, reportUnknownVariableType]
+            if isinstance(y, BetaToolUseBlock)
+        ]
+        tool_call_ids.extend([x.id for x in calls])
+        return tool_call_ids
+
     @override
     async def parse_input(
         self,
@@ -297,77 +321,61 @@ class AnthropicModel(LLM):
         **kwargs: Any,
     ) -> list[dict[str, Any] | Any]:
         new_input: list[dict[str, Any] | Any] = []
+
         content_user: list[dict[str, Any]] = []
 
-        # First pass: collect all tool calls from Message objects for validation
-        tool_calls_in_input: set[str] = set()
-        for item in input:
-            if hasattr(item, "content") and hasattr(item, "role"):
-                content_list = getattr(item, "content", [])
-                for content in content_list:
-                    # Check for both ToolUseBlock and BetaToolUseBlock
-                    if isinstance(content, (ToolUseBlock, BetaToolUseBlock)):
-                        tool_calls_in_input.add(content.id)
+        def flush_content_user():
+            if content_user:
+                # NOTE: must make new object as we clear()
+                new_input.append({"role": "user", "content": content_user.copy()})
+                content_user.clear()
+
+        tool_call_ids = await self.get_tool_call_ids(input)
 
         for item in input:
+            if isinstance(item, TextInput):
+                content_user.append({"type": "text", "text": item.text})
+                continue
+
+            if isinstance(item, FileBase):
+                match item.type:
+                    case "image":
+                        parsed = await self.parse_image(item)
+                    case "file":
+                        parsed = await self.parse_file(item)
+                content_user.append(parsed)
+                continue
+
+            # non content user item
+            flush_content_user()
+
             match item:
-                case TextInput():
-                    if item.text.strip():
-                        content_user.append({"type": "text", "text": item.text})
-                case FileWithBase64() | FileWithUrl() | FileWithId():
-                    match item.type:
-                        case "image":
-                            content_user.append(await self.parse_image(item))
-                        case "file":
-                            content_user.append(await self.parse_file(item))
-                case _:
-                    if content_user:
-                        filtered = filter_empty_text_blocks(content_user)
-                        if filtered:
-                            new_input.append({"role": "user", "content": filtered})
-                        content_user = []
-                    match item:
-                        case ToolResult():
-                            if item.tool_call.id not in tool_calls_in_input:
-                                raise Exception(
-                                    "Tool call result provided with no matching tool call"
-                                )
-                            result_str = normalize_tool_result(item.result)
-                            new_input.append(
+                case ToolResult():
+                    if item.tool_call.id not in tool_call_ids:
+                        raise NoMatchingToolCallError()
+
+                    new_input.append(
+                        {
+                            "role": "user",
+                            "content": [
                                 {
-                                    "role": "user",
-                                    "content": [
-                                        {
-                                            "type": "tool_result",
-                                            "tool_use_id": item.tool_call.id,
-                                            "content": [
-                                                {"type": "text", "text": result_str}
-                                            ],
-                                        }
-                                    ],
+                                    "type": "tool_result",
+                                    "tool_use_id": item.tool_call.id,
+                                    "content": [{"type": "text", "text": item.result}],
                                 }
-                            )
-                        case dict():  # RawInputItem
-                            item = cast(RawInputItem, item)
-                            new_input.append(item)
-                        case _:  # RawResponse
-                            item = cast(Message, item)
-                            filtered_content = [
-                                block
-                                for block in item.content
-                                if not isinstance(block, TextBlock)
-                                or block.text.strip()
-                            ]
-                            if filtered_content:
-                                new_input.append(
-                                    {"role": "assistant", "content": filtered_content}
-                                )
-
-        if content_user:
-            filtered = filter_empty_text_blocks(content_user)
-            if filtered:
-                new_input.append({"role": "user", "content": filtered})
+                            ],
+                        }
+                    )
+                case RawResponse():
+                    content = cast(ParsedBetaMessage, item.response).content
+                    new_input.append({"role": "assistant", "content": content})
+                case RawInput():
+                    new_input.append(item.input)
+
+        # in case content user item is the last item
+        flush_content_user()
 
+        # cache control
         if new_input:
             last_msg = new_input[-1]
             if not isinstance(last_msg, dict):
@@ -485,7 +493,7 @@ class AnthropicModel(LLM):
         bytes: io.BytesIO,
         type: Literal["image", "file"] = "file",
     ) -> FileWithId:
-        file_mime = f"image/{mime}" if type == "image" else mime  # TODO:
+        file_mime = f"image/{mime}" if type == "image" else mime
         response = await self.get_client().beta.files.upload(
             file=(
                 name,
@@ -503,7 +511,8 @@ class AnthropicModel(LLM):
 
     cache_control = {"type": "ephemeral"}  # 5 min cache
 
-    async def create_body(
+    @override
+    async def build_body(
         self,
         input: Sequence[InputItem],
         *,
@@ -555,20 +564,36 @@ class AnthropicModel(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         if self.delegate:
-            return await self.delegate_query(input, tools=tools, **kwargs)
+            return await self.delegate_query(
+                input, tools=tools, query_logger=query_logger, **kwargs
+            )
 
-        body = await self.create_body(input, tools=tools, **kwargs)
+        body = await self.build_body(input, tools=tools, **kwargs)
 
-        betas = ["files-api-2025-04-14", "interleaved-thinking-2025-05-14"]
-        if "sonnet-4-5" in self.model_name:
-            betas.append("context-1m-2025-08-07")
+        client = self.get_client()
 
-        async with self.get_client().beta.messages.stream(
-            **body,
-            betas=betas,
+        # only send betas for the official Anthropic endpoint
+        is_anthropic_endpoint = self._delegate_client is None
+        if not is_anthropic_endpoint:
+            client_base_url = getattr(client, "_base_url", None) or getattr(
+                client, "base_url", None
+            )
+            if client_base_url:
+                is_anthropic_endpoint = "api.anthropic.com" in str(client_base_url)
+
+        stream_kwargs = {**body}
+        if is_anthropic_endpoint:
+            betas = ["files-api-2025-04-14", "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}")
@@ -604,9 +629,37 @@ class AnthropicModel(LLM):
                 cache_write_tokens=message.usage.cache_creation_input_tokens,
             ),
             tool_calls=tool_calls,
-            history=[*input, message],
+            history=[*input, RawResponse(response=message)],
         )
 
+    @override
+    async def count_tokens(
+        self,
+        input: Sequence[InputItem],
+        *,
+        history: Sequence[InputItem] = [],
+        tools: list[ToolDefinition] = [],
+        **kwargs: object,
+    ) -> int:
+        """
+        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
+
+        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)
+
+        client = self.get_client()
+        response = await client.messages.count_tokens(**body)
+
+        return response.input_tokens
+
     @override
     async def _calculate_cost(
         self,

model_library/providers/google/batch.py

@@ -2,8 +2,6 @@ import io
 import json
 from typing import TYPE_CHECKING, Any, Final, Sequence, cast
 
-from typing_extensions import override
-
 from google.genai.types import (
     BatchJob,
     Content,
@@ -11,6 +9,8 @@ from google.genai.types import (
     JobState,
     UploadFileConfig,
 )
+from typing_extensions import override
+
 from model_library.base import BatchResult, InputItem, LLMBatchMixin
 
 if TYPE_CHECKING:
@@ -144,7 +144,7 @@ class GoogleBatchMixin(LLMBatchMixin):
         **kwargs: object,
     ) -> dict[str, Any]:
         self._root.logger.debug(f"Creating batch request for custom_id: {custom_id}")
-        body = await self._root.create_body(input, tools=[], **kwargs)
+        body = await self._root.build_body(input, tools=[], **kwargs)
 
         contents_any = body["contents"]
         serialized_contents: list[dict[str, Any]] = [