model-library 0.1.6__tar.gz → 0.1.7__tar.gz

{model_library-0.1.6 → model_library-0.1.7}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.6
+Version: 0.1.7
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT
@@ -13,13 +13,13 @@ Requires-Dist: pyyaml>=6.0.2
 Requires-Dist: rich
 Requires-Dist: backoff<3.0,>=2.2.1
 Requires-Dist: redis<7.0,>=6.2.0
-Requires-Dist: tiktoken==0.11.0
+Requires-Dist: tiktoken>=0.12.0
 Requires-Dist: pillow
 Requires-Dist: openai<3.0,>=2.0
 Requires-Dist: anthropic<1.0,>=0.57.1
 Requires-Dist: mistralai<2.0,>=1.9.10
 Requires-Dist: xai-sdk<2.0,>=1.0.0
-Requires-Dist: ai21<5.0,>=4.0.3
+Requires-Dist: ai21<5.0,>=4.3.0
 Requires-Dist: boto3<2.0,>=1.38.27
 Requires-Dist: google-genai[aiohttp]>=1.51.0
 Requires-Dist: google-cloud-storage>=1.26.0

{model_library-0.1.6 → model_library-0.1.7}/examples/advanced/web_search.py

@@ -2,6 +2,7 @@ import asyncio
 from typing import Any, cast
 
 from model_library.base import LLM, ToolDefinition
+from model_library.base.output import QueryResult
 from model_library.registry_utils import get_registry_model
 
 from ..setup import console_log, setup
@@ -41,31 +42,7 @@ def print_search_details(tool_call: Any) -> None:
                         console_log(f"    - {source}")
 
 
-def print_citations(response: Any) -> None:
-    """Extract and print citations from response history."""
-    if not response.history:
-        return
-
-    for item in response.history:
-        if not (hasattr(item, "content") and isinstance(item.content, list)):
-            continue
-
-        content_list = cast(list[Any], item.content)
-        for content_item in content_list:
-            if not (hasattr(content_item, "annotations") and content_item.annotations):
-                continue
-
-            console_log("\nCitations:")
-            annotations = cast(list[Any], content_item.annotations)
-            for annotation in annotations:
-                if hasattr(annotation, "url") and annotation.url:
-                    title = getattr(annotation, "title", "Untitled")
-                    url = annotation.url
-                    location = getattr(annotation, "location", "Unknown")
-                    console_log(f"- {title}: {url} (Location: {location})")
-
-
-def print_web_search_results(response: Any) -> None:
+def print_web_search_results(response: QueryResult) -> None:
     """Print comprehensive web search results."""
     console_log(f"Response: {response.output_text}")
 
@@ -74,7 +51,7 @@ def print_web_search_results(response: Any) -> None:
         for tool_call in response.tool_calls:
             print_search_details(tool_call)
 
-    print_citations(response)
+    print(response.extras.citations)
 
 
 async def web_search_domain_filtered(model: LLM) -> None:

model_library-0.1.7/examples/count_tokens.py

@@ -0,0 +1,95 @@
+import asyncio
+import logging
+
+from model_library import set_logging
+from model_library.base import (
+    LLM,
+    QueryResult,
+    TextInput,
+    ToolBody,
+    ToolDefinition,
+)
+from model_library.registry_utils import get_registry_model
+
+from .setup import console_log, setup
+
+
+async def count_tokens(model: LLM):
+    console_log("\n--- Count Tokens ---\n")
+
+    tools = [
+        ToolDefinition(
+            name="get_weather",
+            body=ToolBody(
+                name="get_weather",
+                description="Get current temperature in a given location",
+                properties={
+                    "location": {
+                        "type": "string",
+                        "description": "City and country e.g. Bogotá, Colombia",
+                    },
+                },
+                required=["location"],
+            ),
+        ),
+        ToolDefinition(
+            name="get_danger",
+            body=ToolBody(
+                name="get_danger",
+                description="Get current danger in a given location",
+                properties={
+                    "location": {
+                        "type": "string",
+                        "description": "City and country e.g. Bogotá, Colombia",
+                    },
+                },
+                required=["location"],
+            ),
+        ),
+    ]
+
+    system_prompt = "You must make exactly 0 or 1 tool calls per answer. You must not make more than 1 tool call per answer."
+    user_prompt = "What is the weather in San Francisco right now?"
+
+    predicted_tokens = await model.count_tokens(
+        [TextInput(text=user_prompt)],
+        tools=tools,
+        system_prompt=system_prompt,
+    )
+
+    response: QueryResult = await model.query(
+        [TextInput(text=user_prompt)],
+        tools=tools,
+        system_prompt=system_prompt,
+    )
+
+    actual_tokens = response.metadata.total_input_tokens
+
+    console_log(f"Predicted Token Count: {predicted_tokens}")
+    console_log(f"Actual Token Count: {actual_tokens}\n")
+
+
+async def main():
+    import argparse
+
+    parser = argparse.ArgumentParser(description="Example of counting tokens")
+    parser.add_argument(
+        "model",
+        nargs="?",
+        default="google/gemini-2.5-flash",
+        type=str,
+        help="Model endpoint (default: google/gemini-2.5-flash)",
+    )
+    args = parser.parse_args()
+
+    model = get_registry_model(args.model)
+    model.logger.info(model)
+
+    set_logging(enable=True, level=logging.INFO)
+
+    await count_tokens(model)
+
+
+if __name__ == "__main__":
+    setup()
+    asyncio.run(main())

{model_library-0.1.6 → model_library-0.1.7}/model_library/base/base.py

@@ -13,8 +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 (
@@ -35,6 +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,
@@ -379,6 +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,
@@ -421,6 +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-0.1.6 → model_library-0.1.7}/model_library/base/delegate_only.py

@@ -58,6 +58,16 @@ class DelegateOnly(LLM):
             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(
         self,

{model_library-0.1.6 → model_library-0.1.7}/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-0.1.6 → model_library-0.1.7}/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-0.1.6 → model_library-0.1.7}/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-0.1.6 → model_library-0.1.7}/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-0.1.6 → model_library-0.1.7}/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-0.1.6 → model_library-0.1.7}/model_library/providers/ai21labs.py

@@ -22,6 +22,7 @@ from model_library.base import (
     ToolDefinition,
     ToolResult,
 )
+from model_library.base.input import RawResponse
 from model_library.exceptions import (
     BadInputError,
     MaxOutputTokensExceededError,
@@ -65,8 +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(
@@ -74,7 +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
@@ -133,14 +134,13 @@ class AI21LabsModel(LLM):
         raise NotImplementedError()
 
     @override
-    async def _query_impl(
+    async def build_body(
         self,
         input: Sequence[InputItem],
         *,
         tools: list[ToolDefinition],
-        query_logger: logging.Logger,
         **kwargs: object,
-    ) -> QueryResult:
+    ) -> dict[str, Any]:
         messages: list[ChatMessage] = []
         if "system_prompt" in kwargs:
             messages.append(
@@ -162,6 +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]
@@ -186,7 +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,