model-library 0.1.6__py3-none-any.whl → 0.1.7__py3-none-any.whl

model_library/providers/anthropic.py

@@ -3,15 +3,15 @@ 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,
@@ -22,7 +22,8 @@ from model_library.base import (
     QueryResult,
     QueryResultCost,
     QueryResultMetadata,
-    RawInputItem,
+    RawInput,
+    RawResponse,
     TextInput,
     ToolBody,
     ToolCall,
@@ -31,6 +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
@@ -38,8 +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,
 )
 
 
@@ -62,9 +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,
@@ -300,6 +300,20 @@ class AnthropicModel(LLM):
             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,
@@ -307,77 +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):
@@ -495,7 +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,
@@ -513,7 +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],
         *,
@@ -573,7 +572,7 @@ class AnthropicModel(LLM):
                 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)
 
         client = self.get_client()
 
@@ -630,9 +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]] = [

model_library/providers/google/google.py

@@ -1,13 +1,16 @@
 import base64
 import io
 import logging
+import uuid
 from typing import Any, Literal, Sequence, cast
 
 from google.genai import Client
 from google.genai import errors as genai_errors
 from google.genai.types import (
     Content,
+    CountTokensConfig,
     File,
+    FinishReason,
     FunctionDeclaration,
     GenerateContentConfig,
     GenerateContentResponse,
@@ -21,13 +24,13 @@ from google.genai.types import (
     Tool,
     ToolListUnion,
     UploadFileConfig,
-    FinishReason,
 )
 from typing_extensions import override
 
 from model_library import model_library_settings
 from model_library.base import (
     LLM,
+    FileBase,
     FileInput,
     FileWithBase64,
     FileWithId,
@@ -40,6 +43,8 @@ from model_library.base import (
     QueryResult,
     QueryResultCost,
     QueryResultMetadata,
+    RawInput,
+    RawResponse,
     TextInput,
     ToolBody,
     ToolCall,
@@ -54,8 +59,6 @@ from model_library.exceptions import (
 )
 from model_library.providers.google.batch import GoogleBatchMixin
 from model_library.register_models import register_provider
-from model_library.utils import normalize_tool_result
-import uuid
 
 
 def generate_tool_call_id(tool_name: str) -> str:
@@ -146,63 +149,52 @@ class GoogleModel(LLM):
         input: Sequence[InputItem],
         **kwargs: Any,
     ) -> list[Content]:
-        parsed_input: list[Content] = []
-        parts: list[Part] = []
+        new_input: list[Content] = []
 
-        def flush_parts():
-            nonlocal parts
+        content_user: list[Part] = []
 
-            if parts:
-                parsed_input.append(Content(parts=parts, role="user"))
-                parts = []
+        def flush_content_user():
+            if content_user:
+                new_input.append(Content(parts=content_user, role="user"))
+                content_user.clear()
 
         for item in input:
-            match item:
-                case TextInput():
-                    if item.text.strip():
-                        parts.append(Part.from_text(text=item.text))
+            if isinstance(item, TextInput):
+                content_user.append(Part.from_text(text=item.text))
+                continue
+
+            if isinstance(item, FileBase):
+                parsed = await self.parse_file(item)
+                content_user.append(parsed)
+                continue
 
-                case FileWithBase64() | FileWithUrl() | FileWithId():
-                    part = await self.parse_file(item)
-                    parts.append(part)
+            # non content user item
+            flush_content_user()
 
+            match item:
                 case ToolResult():
-                    flush_parts()
-                    result_str = normalize_tool_result(item.result)
-                    parsed_input.append(
+                    # id check
+                    new_input.append(
                         Content(
                             role="function",
                             parts=[
                                 Part.from_function_response(
                                     name=item.tool_call.name,
-                                    response={"result": result_str},
+                                    response={"result": item.result},
                                 )
                             ],
                         )
                     )
 
-                case GenerateContentResponse():
-                    flush_parts()
-                    candidates = item.candidates
-                    if candidates and candidates[0]:
-                        content0 = candidates[0].content
-                        if content0 is not None:
-                            parsed_input.append(content0)
-                    else:
-                        self.logger.debug(
-                            "GenerateContentResponse missing candidates; skipping"
-                        )
-
-                case Content():
-                    flush_parts()
-                    parsed_input.append(item)
+                case RawResponse():
+                    new_input.extend(item.response)
+                case RawInput():
+                    new_input.append(item.input)
 
-                case _:
-                    raise BadInputError(f"Unsupported input type: {type(item)}")
+        # in case content user item is the last item
+        flush_content_user()
 
-        flush_parts()
-
-        return parsed_input
+        return new_input
 
     @override
     async def parse_file(self, file: FileInput) -> Part:
@@ -284,7 +276,8 @@ class GoogleModel(LLM):
             mime=mime,
         )
 
-    async def create_body(
+    @override
+    async def build_body(
         self,
         input: Sequence[InputItem],
         *,
@@ -337,7 +330,7 @@ class GoogleModel(LLM):
         query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
-        body: dict[str, Any] = await self.create_body(input, tools=tools, **kwargs)
+        body: dict[str, Any] = await self.build_body(input, tools=tools, **kwargs)
 
         text: str = ""
         reasoning: str = ""
@@ -395,7 +388,7 @@ class GoogleModel(LLM):
         result = QueryResult(
             output_text=text,
             reasoning=reasoning,
-            history=[*input, *contents],
+            history=[*input, RawResponse(response=contents)],
             tool_calls=tool_calls,
         )
 
@@ -410,6 +403,51 @@ class GoogleModel(LLM):
             )
         return result
 
+    @override
+    async def count_tokens(
+        self,
+        input: Sequence[InputItem],
+        *,
+        history: Sequence[InputItem] = [],
+        tools: list[ToolDefinition] = [],
+        **kwargs: object,
+    ) -> int:
+        """
+        Count the number of tokens using Google's native token counting API.
+        https://ai.google.dev/gemini-api/docs/tokens
+
+        Only Vertex AI supports system_instruction and tools in count_tokens.
+        For Gemini API, fall back to the base implementation.
+        TODO: implement token counting for non-Vertex models.
+        """
+        if not self.provider_config.use_vertex:
+            return await super().count_tokens(
+                input, history=history, tools=tools, **kwargs
+            )
+
+        input = [*history, *input]
+        if not input:
+            return 0
+
+        system_prompt = kwargs.pop("system_prompt", None)
+        contents = await self.parse_input(input, **kwargs)
+        parsed_tools = await self.parse_tools(tools) if tools else None
+        config = CountTokensConfig(
+            system_instruction=str(system_prompt) if system_prompt else None,
+            tools=parsed_tools,
+        )
+
+        response = await self.client.aio.models.count_tokens(
+            model=self.model_name,
+            contents=cast(Any, contents),
+            config=config,
+        )
+
+        if response.total_tokens is None:
+            raise ValueError("count_tokens returned None")
+
+        return response.total_tokens
+
     @override
     async def _calculate_cost(
         self,
@@ -446,7 +484,7 @@ class GoogleModel(LLM):
         **kwargs: object,
     ) -> PydanticT:
         # Create the request body with JSON schema
-        body: dict[str, Any] = await self.create_body(input, tools=[], **kwargs)
+        body: dict[str, Any] = await self.build_body(input, tools=[], **kwargs)
 
         # Get the JSON schema from the Pydantic model
         json_schema = pydantic_model.model_json_schema()

model_library/providers/minimax.py

@@ -2,12 +2,16 @@ from typing import Literal
 
 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.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")
 class MinimaxModel(DelegateOnly):
@@ -31,3 +35,18 @@ class MinimaxModel(DelegateOnly):
                 max_retries=1,
             ),
         )
+
+    # minimax client shares anthropic's syntax
+    @override
+    async def count_tokens(
+        self,
+        input: Sequence[InputItem],
+        *,
+        history: Sequence[InputItem] = [],
+        tools: list[ToolDefinition] = [],
+        **kwargs: object,
+    ) -> int:
+        assert self.delegate
+        return await self.delegate.count_tokens(
+            input, history=history, tools=tools, **kwargs
+        )