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

model_library/base/base.py

@@ -6,7 +6,6 @@ from abc import ABC, abstractmethod
 from collections.abc import Awaitable
 from pprint import pformat
 from typing import (
-    TYPE_CHECKING,
     Any,
     Callable,
     Literal,
@@ -14,8 +13,10 @@ from typing import (
     TypeVar,
 )
 
+import tiktoken
 from pydantic import model_serializer
 from pydantic.main import BaseModel
+from tiktoken.core import Encoding
 from typing_extensions import override
 
 from model_library.base.batch import (
@@ -36,6 +37,7 @@ from model_library.base.output import (
 )
 from model_library.base.utils import (
     get_pretty_input_types,
+    serialize_for_tokenizing,
 )
 from model_library.exceptions import (
     ImmediateRetryException,
@@ -43,9 +45,6 @@ from model_library.exceptions import (
 )
 from model_library.utils import truncate_str
 
-if TYPE_CHECKING:
-    from model_library.providers.openai import OpenAIModel
-
 PydanticT = TypeVar("PydanticT", bound=BaseModel)
 
 
@@ -66,7 +65,7 @@ class LLMConfig(BaseModel):
     top_p: float | None = None
     top_k: int | None = None
     reasoning: bool = False
-    reasoning_effort: str | None = None
+    reasoning_effort: str | bool | None = None
     supports_images: bool = False
     supports_files: bool = False
     supports_videos: bool = False
@@ -110,7 +109,7 @@ class LLM(ABC):
         self.top_k: int | None = config.top_k
 
         self.reasoning: bool = config.reasoning
-        self.reasoning_effort: str | None = config.reasoning_effort
+        self.reasoning_effort: str | bool | None = config.reasoning_effort
 
         self.supports_files: bool = config.supports_files
         self.supports_videos: bool = config.supports_videos
@@ -120,7 +119,7 @@ class LLM(ABC):
         self.supports_tools: bool = config.supports_tools
 
         self.native: bool = config.native
-        self.delegate: "OpenAIModel | None" = None
+        self.delegate: "LLM | None" = None
         self.batch: LLMBatchMixin | None = None
 
         if config.provider_config:
@@ -198,11 +197,14 @@ class LLM(ABC):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition] = [],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         if not self.delegate:
             raise Exception("Delegate not set")
-        return await self.delegate._query_impl(input, tools=tools, **kwargs)  # pyright: ignore[reportPrivateUsage]
+        return await self.delegate._query_impl(  # pyright: ignore[reportPrivateUsage]
+            input, tools=tools, query_logger=query_logger, **kwargs
+        )
 
     async def query(
         self,
@@ -213,6 +215,7 @@ class LLM(ABC):
         # for backwards compatibility
         files: list[FileInput] = [],
         images: list[FileInput] = [],
+        query_logger: logging.Logger | None = None,
         **kwargs: object,
     ) -> QueryResult:
         """
@@ -256,15 +259,18 @@ class LLM(ABC):
         input = [*history, *input]
 
         # unique logger for the query
-        query_id = uuid.uuid4().hex[:14]
-        query_logger = self.logger.getChild(f"query={query_id}")
+        if not query_logger:
+            query_id = uuid.uuid4().hex[:14]
+            query_logger = self.logger.getChild(f"query={query_id}")
 
         query_logger.info(
             "Query started:\n" + item_info + tool_info + f"--- kwargs: {short_kwargs}\n"
         )
 
         async def query_func() -> QueryResult:
-            return await self._query_impl(input, tools=tools, **kwargs)
+            return await self._query_impl(
+                input, tools=tools, query_logger=query_logger, **kwargs
+            )
 
         async def timed_query() -> tuple[QueryResult, float]:
             return await LLM.timer_wrapper(query_func)
@@ -361,7 +367,8 @@ class LLM(ABC):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
-        **kwargs: object,  # TODO: pass in query logger
+        query_logger: logging.Logger,
+        **kwargs: object,
     ) -> QueryResult:
         """
         Query the model with input
@@ -375,6 +382,20 @@ class LLM(ABC):
         """
         ...
 
+    @abstractmethod
+    async def build_body(
+        self,
+        input: Sequence[InputItem],
+        *,
+        tools: list[ToolDefinition],
+        **kwargs: Any,
+    ) -> dict[str, Any]:
+        """
+        Builds the body of the request to the model provider
+        Calls parse_input
+        """
+        ...
+
     @abstractmethod
     async def parse_input(
         self,
@@ -417,6 +438,87 @@ class LLM(ABC):
         """Upload a file to the model provider"""
         ...
 
+    async def get_encoding(self) -> Encoding:
+        """Get the appropriate tokenizer"""
+
+        model = self.model_name.lower()
+
+        if any(x in model for x in ["gpt-4o", "o1", "o3", "gpt-4.1", "gpt-5"]):
+            return tiktoken.get_encoding("o200k_base")
+        elif "gpt-4" in model or "gpt-3.5" in model:
+            try:
+                return tiktoken.encoding_for_model(self.model_name)
+            except KeyError:
+                return tiktoken.get_encoding("cl100k_base")
+        elif "claude" in model:
+            return tiktoken.get_encoding("cl100k_base")
+        elif "gemini" in model:
+            return tiktoken.get_encoding("o200k_base")
+        elif "llama" in model or "mistral" in model:
+            return tiktoken.get_encoding("cl100k_base")
+        else:
+            return tiktoken.get_encoding("cl100k_base")
+
+    async def stringify_input(
+        self,
+        input: Sequence[InputItem],
+        *,
+        history: Sequence[InputItem] = [],
+        tools: list[ToolDefinition] = [],
+        **kwargs: object,
+    ) -> str:
+        input = [*history, *input]
+
+        system_prompt = kwargs.pop(
+            "system_prompt", ""
+        )  # TODO: refactor along with system prompt arg change
+
+        # special case if using a delegate
+        # don't inherit method override by default
+        if self.delegate:
+            parsed_input = await self.delegate.parse_input(input, **kwargs)
+            parsed_tools = await self.delegate.parse_tools(tools)
+        else:
+            parsed_input = await self.parse_input(input, **kwargs)
+            parsed_tools = await self.parse_tools(tools)
+
+        serialized_input = serialize_for_tokenizing(parsed_input)
+        serialized_tools = serialize_for_tokenizing(parsed_tools)
+
+        combined = f"{system_prompt}\n{serialized_input}\n{serialized_tools}"
+
+        return combined
+
+    async def count_tokens(
+        self,
+        input: Sequence[InputItem],
+        *,
+        history: Sequence[InputItem] = [],
+        tools: list[ToolDefinition] = [],
+        **kwargs: object,
+    ) -> int:
+        """
+        Count the number of tokens for a query.
+        Combines parsed input and tools, then tokenizes the result.
+        """
+
+        if not input and not history:
+            return 0
+
+        if self.delegate:
+            encoding = await self.delegate.get_encoding()
+        else:
+            encoding = await self.get_encoding()
+        self.logger.debug(f"Token Count Encoding: {encoding}")
+
+        string_input = await self.stringify_input(
+            input, history=history, tools=tools, **kwargs
+        )
+
+        count = len(encoding.encode(string_input, disallowed_special=()))
+        self.logger.debug(f"Combined Token Count Input: {count}")
+        return count
+
     async def query_json(
         self,
         input: Sequence[InputItem],

model_library/base/delegate_only.py

@@ -1,4 +1,5 @@
 import io
+import logging
 from typing import Any, Literal, Sequence
 
 from typing_extensions import override
@@ -48,11 +49,24 @@ class DelegateOnly(LLM):
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
+        query_logger: logging.Logger,
         **kwargs: object,
     ) -> QueryResult:
         assert self.delegate
 
-        return await self.delegate_query(input, tools=tools, **kwargs)
+        return await self.delegate_query(
+            input, tools=tools, query_logger=query_logger, **kwargs
+        )
+
+    @override
+    async def build_body(
+        self,
+        input: Sequence[InputItem],
+        *,
+        tools: list[ToolDefinition],
+        **kwargs: object,
+    ) -> dict[str, Any]:
+        raise DelegateOnlyException()
 
     @override
     async def parse_input(

model_library/base/input.py

@@ -74,8 +74,6 @@ class ToolCall(BaseModel):
 --- INPUT ---
 """
 
-RawResponse = Any
-
 
 class ToolInput(BaseModel):
     tools: list[ToolDefinition] = []
@@ -90,11 +88,16 @@ class TextInput(BaseModel):
     text: str
 
 
-RawInputItem = dict[
-    str, Any
-]  # to pass in, for example, a mock convertsation with {"role": "user", "content": "Hello"}
+class RawResponse(BaseModel):
+    # used to store a received response
+    response: Any
+
+
+class RawInput(BaseModel):
+    # used to pass in anything provider specific (e.g. a mock conversation)
+    input: Any
 
 
 InputItem = (
-    TextInput | FileInput | ToolResult | RawInputItem | RawResponse
-)  # input item can either be a prompt, a file (image or file), a tool call result, raw input, or a previous response
+    TextInput | FileInput | ToolResult | RawInput | RawResponse
+)  # input item can either be a prompt, a file (image or file), a tool call result, a previous response, or raw input

model_library/base/output.py

@@ -24,6 +24,11 @@ class Citation(BaseModel):
     index: int | None = None
     container_id: str | None = None
 
+    @override
+    def __repr__(self):
+        attrs = vars(self).copy()
+        return f"{self.__class__.__name__}(\n{pformat(attrs, indent=2)}\n)"
+
 
 class QueryResultExtras(BaseModel):
     citations: list[Citation] = Field(default_factory=list)

model_library/base/utils.py

@@ -1,18 +1,34 @@
-from typing import Sequence, TypeVar, cast
+import json
+from typing import Any, Sequence, TypeVar
+
+from pydantic import BaseModel
 
 from model_library.base.input import (
     FileBase,
     InputItem,
-    RawInputItem,
+    RawInput,
+    RawResponse,
     TextInput,
     ToolResult,
 )
 from model_library.utils import truncate_str
-from pydantic import BaseModel
 
 T = TypeVar("T", bound=BaseModel)
 
 
+def serialize_for_tokenizing(content: Any) -> str:
+    """
+    Serialize parsed content into a string for tokenization
+    """
+    parts: list[str] = []
+    if content:
+        if isinstance(content, str):
+            parts.append(content)
+        else:
+            parts.append(json.dumps(content, default=str))
+    return "\n".join(parts)
+
+
 def add_optional(
     a: int | float | T | None, b: int | float | T | None
 ) -> int | float | T | None:
@@ -54,11 +70,9 @@ def get_pretty_input_types(input: Sequence["InputItem"], verbose: bool = False)
                 return repr(item)
             case ToolResult():
                 return repr(item)
-            case dict():
-                item = cast(RawInputItem, item)
+            case RawInput():
                 return repr(item)
-            case _:
-                # RawResponse
+            case RawResponse():
                 return repr(item)
 
     processed_items = [f"  {process_item(item)}" for item in input]

model_library/config/all_models.json

@@ -1,4 +1,94 @@
 {
+    "minimax/MiniMax-M2.1": {
+        "company": "MiniMax",
+        "label": "MiniMax-M2.1",
+        "description": null,
+        "release_date": "2025-12-23",
+        "open_source": true,
+        "documentation_url": "https://platform.minimax.io/docs",
+        "properties": {
+            "context_window": 204800,
+            "max_tokens": 131000,
+            "training_cutoff": null,
+            "reasoning_model": true
+        },
+        "supports": {
+            "images": false,
+            "files": false,
+            "temperature": true,
+            "tools": true
+        },
+        "metadata": {
+            "deprecated": false,
+            "available_for_everyone": true,
+            "available_as_evaluator": false,
+            "ignored_for_cost": false
+        },
+        "provider_properties": {},
+        "costs_per_million_token": {
+            "input": 0.3,
+            "output": 1.2,
+            "cache": {
+                "read": 0.03,
+                "write": 0.375,
+                "write_markup": 1.0
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 1.0,
+            "top_p": 0.95
+        },
+        "provider_endpoint": "MiniMax-M2.1",
+        "provider_name": "minimax",
+        "full_key": "minimax/MiniMax-M2.1",
+        "slug": "minimax_MiniMax-M2.1"
+    },
+    "zai/glm-4.7": {
+        "company": "zAI",
+        "label": "GLM 4.7",
+        "description": "Latest model from ZAI",
+        "release_date": "2025-12-22",
+        "open_source": true,
+        "documentation_url": "https://docs.z.ai/",
+        "properties": {
+            "context_window": 200000,
+            "max_tokens": 128000,
+            "training_cutoff": null,
+            "reasoning_model": true
+        },
+        "supports": {
+            "images": false,
+            "files": false,
+            "temperature": true,
+            "tools": true
+        },
+        "metadata": {
+            "deprecated": false,
+            "available_for_everyone": true,
+            "available_as_evaluator": false,
+            "ignored_for_cost": false
+        },
+        "provider_properties": {},
+        "costs_per_million_token": {
+            "input": 0.6,
+            "output": 2.2,
+            "cache": {
+                "read": 0.11,
+                "read_discount": 1.0,
+                "write_markup": 1.0
+            }
+        },
+        "alternative_keys": [],
+        "default_parameters": {
+            "temperature": 1.0,
+            "top_p": 1.0
+        },
+        "provider_endpoint": "glm-4.7",
+        "provider_name": "zai",
+        "full_key": "zai/glm-4.7",
+        "slug": "zai_glm-4.7"
+    },
     "google/gemini-3-flash-preview": {
         "company": "Google",
         "label": "Gemini 3 Flash (12/25)",
@@ -504,7 +594,8 @@
             }
         ],
         "default_parameters": {
-            "temperature": 1.0
+            "temperature": 1.0,
+            "reasoning_effort": "none"
         },
         "provider_endpoint": "deepseek-v3p2",
         "provider_name": "fireworks",

model_library/config/fireworks_models.yaml

@@ -150,6 +150,8 @@ deepseek-models:
       context_window: 160_000
       max_tokens: 20_480
       reasoning_model: false
+    default_parameters:
+      reasoning_effort: "none"
     costs_per_million_token:
       input: 0.56
       output: 1.68

model_library/config/minimax_models.yaml

@@ -16,6 +16,24 @@ base-config:
 
 minimax-m2-models:
 
+  minimax/MiniMax-M2.1:
+    label: MiniMax-M2.1
+    release_date: 2025-12-23
+    properties:
+      context_window: 204_800
+      max_tokens: 131_000
+      reasoning_model: true
+      training_cutoff: null
+    default_parameters:
+      temperature: 1.0
+      top_p: 0.95
+    costs_per_million_token:
+      input: 0.30
+      output: 1.20
+      cache:
+        read: 0.03
+        write: 0.375
+
   minimax/MiniMax-M2:
     label: MiniMax-M2
     description: MiniMax-M2 is a cost-efficient open-source model optimized for agentic applications and coding in particular.

model_library/config/zai_models.yaml

@@ -18,6 +18,20 @@ base-config:
       write_markup: 1
 
 zai-models:
+  zai/glm-4.7:
+    label: GLM 4.7
+    description: "Latest model from ZAI"
+    release_date: 2025-12-22
+    properties:
+      context_window: 200_000
+      max_tokens: 128_000
+    costs_per_million_token:
+      input: 0.6
+      output: 2.2
+      cache:
+        read: 0.11
+    default_parameters:
+      temperature: 1
   zai/glm-4.5:
     label: GLM 4.5
     description: "z.AI old model"

model_library/exceptions.py

@@ -146,6 +146,17 @@ class BadInputError(Exception):
         super().__init__(message or BadInputError.DEFAULT_MESSAGE)
 
 
+class NoMatchingToolCallError(Exception):
+    """
+    Raised when a tool call result is provided with no matching tool call
+    """
+
+    DEFAULT_MESSAGE: str = "Tool call result provided with no matching tool call"
+
+    def __init__(self, message: str | None = None):
+        super().__init__(message or NoMatchingToolCallError.DEFAULT_MESSAGE)
+
+
 # Add more retriable exceptions as needed
 # Providers that don't have an explicit rate limit error are handled manually
 # by wrapping errored Http/gRPC requests with a BackoffRetryException

model_library/logging.py

@@ -6,7 +6,11 @@ from rich.logging import RichHandler
 _llm_logger = logging.getLogger("llm")
 
 
-def set_logging(enable: bool = True, handler: logging.Handler | None = None):
+def set_logging(
+    enable: bool = True,
+    level: int = logging.INFO,
+    handler: logging.Handler | None = None,
+):
     """
     Sets up logging for the model library
 
@@ -15,7 +19,7 @@ def set_logging(enable: bool = True, handler: logging.Handler | None = None):
         handler (logging.Handler, optional): A custom logging handler. Defaults to RichHandler.
     """
     if enable:
-        _llm_logger.setLevel(logging.INFO)
+        _llm_logger.setLevel(level)
     else:
         _llm_logger.setLevel(logging.CRITICAL)
 

model_library/providers/ai21labs.py

@@ -1,4 +1,5 @@
 import io
+import logging
 from typing import Any, Literal, Sequence
 
 from ai21 import AsyncAI21Client
@@ -21,6 +22,7 @@ from model_library.base import (
     ToolDefinition,
     ToolResult,
 )
+from model_library.base.input import RawResponse
 from model_library.exceptions import (
     BadInputError,
     MaxOutputTokensExceededError,
@@ -64,8 +66,6 @@ class AI21LabsModel(LLM):
             match item:
                 case TextInput():
                     new_input.append(ChatMessage(role="user", content=item.text))
-                case AssistantMessage():
-                    new_input.append(item)
                 case ToolResult():
                     new_input.append(
                         ToolMessage(
@@ -73,7 +73,9 @@ class AI21LabsModel(LLM):
                             content=item.result,
                             tool_call_id=item.tool_call.id,
                         )
-                    )
+                    )  # TODO: tool calling metadata and test
+                case RawResponse():
+                    new_input.append(item.response)
                 case _:
                     raise BadInputError("Unsupported input type")
         return new_input
@@ -132,13 +134,13 @@ class AI21LabsModel(LLM):
         raise NotImplementedError()
 
     @override
-    async def _query_impl(
+    async def build_body(
         self,
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
         **kwargs: object,
-    ) -> QueryResult:
+    ) -> dict[str, Any]:
         messages: list[ChatMessage] = []
         if "system_prompt" in kwargs:
             messages.append(
@@ -160,6 +162,18 @@ class AI21LabsModel(LLM):
                 body["top_p"] = self.top_p
 
         body.update(kwargs)
+        return body
+
+    @override
+    async def _query_impl(
+        self,
+        input: Sequence[InputItem],
+        *,
+        tools: list[ToolDefinition],
+        query_logger: logging.Logger,
+        **kwargs: object,
+    ) -> QueryResult:
+        body = await self.build_body(input, tools=tools, **kwargs)
 
         response: ChatCompletionResponse = (
             await self.get_client().chat.completions.create(**body, stream=False)  # pyright: ignore[reportAny, reportUnknownMemberType]
@@ -184,7 +198,7 @@ class AI21LabsModel(LLM):
 
         output = QueryResult(
             output_text=choice.message.content,
-            history=[*input, choice.message],
+            history=[*input, RawResponse(response=choice.message)],
             metadata=QueryResultMetadata(
                 in_tokens=response.usage.prompt_tokens,
                 out_tokens=response.usage.completion_tokens,