model-library 0.1.1__tar.gz → 0.1.3__tar.gz

{model_library-0.1.1 → model_library-0.1.3}/Makefile

@@ -86,5 +86,5 @@ examples: venv_check
 	@echo "\n✅ All examples completed!"
 	
 browse_models: venv_check
-	@uv run python -m scripts.explore_models
+	@uv run python -m scripts.browse_models
 

{model_library-0.1.1 → model_library-0.1.3}/PKG-INFO

@@ -1,6 +1,6 @@
 Metadata-Version: 2.4
 Name: model-library
-Version: 0.1.1
+Version: 0.1.3
 Summary: Model Library for vals.ai 
 Author-email: "Vals AI, Inc." <contact@vals.ai>
 License: MIT
@@ -17,12 +17,11 @@ Requires-Dist: tiktoken==0.11.0
 Requires-Dist: pillow
 Requires-Dist: openai<2.0,>=1.97.1
 Requires-Dist: anthropic<1.0,>=0.57.1
-Requires-Dist: together<2.0,>=1.5.25
 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: boto3<2.0,>=1.38.27
-Requires-Dist: google-genai[aiohttp]<2.0,>=1.48.0
+Requires-Dist: google-genai[aiohttp]>=1.51.0
 Requires-Dist: google-cloud-storage>=1.26.0
 Dynamic: license-file
 

model_library-0.1.3/model_library/base/__init__.py

@@ -0,0 +1,7 @@
+# ruff: noqa: F403,F401
+from model_library.base.base import *
+from model_library.base.batch import *
+from model_library.base.delegate_only import *
+from model_library.base.input import *
+from model_library.base.output import *
+from model_library.base.utils import *

{model_library-0.1.1/model_library → model_library-0.1.3/model_library/base}/base.py

@@ -7,293 +7,46 @@ from collections.abc import Awaitable
 from pprint import pformat
 from typing import (
     TYPE_CHECKING,
-    Annotated,
     Any,
     Callable,
     Literal,
-    Mapping,
     Sequence,
     TypeVar,
-    cast,
 )
 
-from pydantic import computed_field, field_validator, model_serializer
-from pydantic.fields import Field
+from pydantic import model_serializer
 from pydantic.main import BaseModel
 from typing_extensions import override
 
+from model_library.base.batch import (
+    LLMBatchMixin,
+)
+from model_library.base.input import (
+    FileInput,
+    FileWithId,
+    InputItem,
+    TextInput,
+    ToolDefinition,
+    ToolResult,
+)
+from model_library.base.output import (
+    QueryResult,
+    QueryResultCost,
+    QueryResultMetadata,
+)
+from model_library.base.utils import (
+    get_pretty_input_types,
+)
 from model_library.exceptions import (
     ImmediateRetryException,
     retry_llm_call,
 )
-from model_library.utils import sum_optional, truncate_str
-
-PydanticT = TypeVar("PydanticT", bound=BaseModel)
-
-DEFAULT_MAX_TOKENS = 2048
-DEFAULT_TEMPERATURE = 0.7
-DEFAULT_TOP_P = 1
+from model_library.utils import truncate_str
 
 if TYPE_CHECKING:
     from model_library.providers.openai import OpenAIModel
 
-"""
---- FILES ---
-"""
-
-
-class FileBase(BaseModel):
-    type: Literal["image", "file"]
-    name: str
-    mime: str
-
-    @override
-    def __repr__(self):
-        attrs = vars(self).copy()
-        if "base64" in attrs:
-            attrs["base64"] = truncate_str(attrs["base64"])
-        return f"{self.__class__.__name__}(\n{pformat(attrs, indent=2)}\n)"
-
-
-class FileWithBase64(FileBase):
-    append_type: Literal["base64"] = "base64"
-    base64: str
-
-
-class FileWithUrl(FileBase):
-    append_type: Literal["url"] = "url"
-    url: str
-
-
-class FileWithId(FileBase):
-    append_type: Literal["file_id"] = "file_id"
-    file_id: str
-
-
-FileInput = Annotated[
-    FileWithBase64 | FileWithUrl | FileWithId,
-    Field(discriminator="append_type"),
-]
-
-
-"""
---- TOOLS ---
-"""
-
-
-class ToolBody(BaseModel):
-    name: str
-    description: str
-    properties: dict[str, Any]
-    required: list[str]
-    kwargs: dict[str, Any] = {}
-
-
-class ToolDefinition(BaseModel):
-    name: str  # acts as a key
-    body: ToolBody | Any
-
-
-class ToolCall(BaseModel):
-    id: str
-    call_id: str | None = None
-    name: str
-    args: dict[str, Any] | str
-
-
-"""
---- INPUT ---
-"""
-
-RawResponse = Any
-
-
-class ToolInput(BaseModel):
-    tools: list[ToolDefinition] = []
-
-
-class ToolResult(BaseModel):
-    tool_call: ToolCall
-    result: Any
-
-
-class TextInput(BaseModel):
-    text: str
-
-
-RawInputItem = dict[
-    str, Any
-]  # to pass in, for example, a mock convertsation with {"role": "user", "content": "Hello"}
-
-
-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
-
-
-"""
---- OUTPUT ---
-"""
-
-
-class Citation(BaseModel):
-    type: str | None = None
-    title: str | None = None
-    url: str | None = None
-    start_index: int | None = None
-    end_index: int | None = None
-    file_id: str | None = None
-    filename: str | None = None
-    index: int | None = None
-    container_id: str | None = None
-
-
-class QueryResultExtras(BaseModel):
-    citations: list[Citation] = Field(default_factory=list)
-
-
-class QueryResultCost(BaseModel):
-    """
-    Cost information for a query
-    Includes total cost and a structured breakdown.
-    """
-
-    input: float
-    output: float
-    reasoning: float | None = None
-    cache_read: float | None = None
-    cache_write: float | None = None
-
-    @computed_field
-    @property
-    def total(self) -> float:
-        return sum(
-            filter(
-                None,
-                [
-                    self.input,
-                    self.output,
-                    self.reasoning,
-                    self.cache_read,
-                    self.cache_write,
-                ],
-            )
-        )
-
-    @override
-    def __repr__(self):
-        use_cents = self.total < 1
-
-        def format_cost(value: float | None):
-            if value is None:
-                return None
-            return f"{value * 100:.3f} cents" if use_cents else f"${value:.2f}"
-
-        return (
-            f"{format_cost(self.total)} "
-            + f"(uncached input: {format_cost(self.input)} | output: {format_cost(self.output)} | reasoning: {format_cost(self.reasoning)} | cache_read: {format_cost(self.cache_read)} | cache_write: {format_cost(self.cache_write)})"
-        )
-
-
-class QueryResultMetadata(BaseModel):
-    """
-    Metadata for a query: token usage and timing.
-
-    """
-
-    cost: QueryResultCost | None = None  # set post query
-    duration_seconds: float | None = None  # set post query
-    in_tokens: int = 0
-    out_tokens: int = 0
-    reasoning_tokens: int | None = None
-    cache_read_tokens: int | None = None
-    cache_write_tokens: int | None = None
-
-    @property
-    def default_duration_seconds(self) -> float:
-        return self.duration_seconds or 0
-
-    def __add__(self, other: "QueryResultMetadata") -> "QueryResultMetadata":
-        return QueryResultMetadata(
-            in_tokens=self.in_tokens + other.in_tokens,
-            out_tokens=self.out_tokens + other.out_tokens,
-            reasoning_tokens=sum_optional(
-                self.reasoning_tokens, other.reasoning_tokens
-            ),
-            cache_read_tokens=sum_optional(
-                self.cache_read_tokens, other.cache_read_tokens
-            ),
-            cache_write_tokens=sum_optional(
-                self.cache_write_tokens, other.cache_write_tokens
-            ),
-            duration_seconds=self.default_duration_seconds
-            + other.default_duration_seconds,
-        )
-
-    @override
-    def __repr__(self):
-        attrs = vars(self).copy()
-        return f"{self.__class__.__name__}(\n{pformat(attrs, indent=2, sort_dicts=False)}\n)"
-
-
-class QueryResult(BaseModel):
-    """
-    Result of a query
-    Contains the text, reasoning, metadata, tool calls, and history
-    """
-
-    output_text: str | None = None
-    reasoning: str | None = None
-    metadata: QueryResultMetadata = Field(default_factory=QueryResultMetadata)
-    tool_calls: list[ToolCall] = Field(default_factory=list)
-    history: list[InputItem] = Field(default_factory=list)
-    extras: QueryResultExtras = Field(default_factory=QueryResultExtras)
-    raw: dict[str, Any] = Field(default_factory=dict)
-
-    @property
-    def output_text_str(self) -> str:
-        return self.output_text or ""
-
-    @field_validator("reasoning", mode="before")
-    def default_reasoning(cls, v: str | None):
-        return None if not v else v  # make reasoning None if empty
-
-    @property
-    def search_results(self) -> Any | None:
-        """Expose provider-supplied search metadata without additional processing."""
-        raw_dict = cast(dict[str, Any], getattr(self, "raw", {}))
-        raw_candidate = raw_dict.get("search_results")
-        if raw_candidate is not None:
-            return raw_candidate
-
-        return _get_from_history(self.history, "search_results")
-
-    @override
-    def __repr__(self):
-        attrs = vars(self).copy()
-        ordered_attrs = {
-            "output_text": truncate_str(attrs.pop("output_text", None), 400),
-            "reasoning": truncate_str(attrs.pop("reasoning", None), 400),
-            "metadata": attrs.pop("metadata", None),
-        }
-        if self.tool_calls:
-            ordered_attrs["tool_calls"] = self.tool_calls
-        return f"{self.__class__.__name__}(\n{pformat(ordered_attrs, indent=2, sort_dicts=False)}\n)"
-
-
-def _get_from_history(history: Sequence[InputItem], key: str) -> Any | None:
-    for item in reversed(history):
-        value = getattr(item, key, None)
-        if value is not None:
-            return value
-
-        extra = getattr(item, "model_extra", None)
-        if isinstance(extra, Mapping):
-            value = cast(Mapping[str, Any], extra).get(key)
-            if value is not None:
-                return value
-
-    return None
+PydanticT = TypeVar("PydanticT", bound=BaseModel)
 
 
 class ProviderConfig(BaseModel):
@@ -304,6 +57,9 @@ class ProviderConfig(BaseModel):
         return self.__dict__
 
 
+DEFAULT_MAX_TOKENS = 2048
+
+
 class LLMConfig(BaseModel):
     max_tokens: int = DEFAULT_MAX_TOKENS
     temperature: float | None = None
@@ -347,9 +103,6 @@ class LLM(ABC):
         config = config or LLMConfig()
         self._registry_key = config.registry_key
 
-        if config.provider_config:
-            self.provider_config = config.provider_config
-
         self.max_tokens: int = config.max_tokens
         self.temperature: float | None = config.temperature
         self.top_p: float | None = config.top_p
@@ -368,6 +121,12 @@ class LLM(ABC):
         self.delegate: "OpenAIModel | None" = None
         self.batch: LLMBatchMixin | None = None
 
+        if config.provider_config:
+            if isinstance(
+                config.provider_config, type(getattr(self, "provider_config"))
+            ):
+                self.provider_config = config.provider_config
+
         self.logger: logging.Logger = logging.getLogger(
             f"llm.{provider}.{model_name}<instance={self.instance_id}>"
         )
@@ -459,6 +218,10 @@ class LLM(ABC):
         Join input with history
         Log, Time, and Retry
         """
+
+        # verbose on debug
+        verbose = self.logger.isEnabledFor(logging.DEBUG)
+
         # format str input
         if isinstance(input, str):
             input = [TextInput(text=input)]
@@ -467,11 +230,11 @@ class LLM(ABC):
         input = [*files, *images, *input]
 
         # format input info
-        item_info = f"--- input ({len(input)}): {get_pretty_input_types(input)}\n"
+        item_info = (
+            f"--- input ({len(input)}): {get_pretty_input_types(input, verbose)}\n"
+        )
         if history:
-            item_info += (
-                f"--- history({len(history)}): {get_pretty_input_types(history)}\n"
-            )
+            item_info += f"--- history({len(history)}): {get_pretty_input_types(history, verbose)}\n"
 
         # format tool info
         tool_results = [t for t in input if isinstance(t, ToolResult)]
@@ -492,7 +255,7 @@ class LLM(ABC):
 
         # unique logger for the query
         query_id = uuid.uuid4().hex[:14]
-        query_logger = logging.getLogger(f"{self.logger.name}<query={query_id}>")
+        query_logger = self.logger.getChild(f"query={query_id}")
 
         query_logger.info(
             "Query started:\n" + item_info + tool_info + f"--- kwargs: {short_kwargs}\n"
@@ -518,36 +281,10 @@ class LLM(ABC):
         output.metadata.cost = await self._calculate_cost(output.metadata)
 
         query_logger.info(f"Query completed: {repr(output)}")
+        query_logger.debug(output.model_dump(exclude={"history", "raw"}))
 
         return output
 
-    async def query_json(
-        self,
-        input: Sequence[InputItem],
-        pydantic_model: type[PydanticT],
-        **kwargs: object,
-    ) -> PydanticT:
-        """Query the model with JSON response format using Pydantic model.
-
-        This is a convenience method that is not implemented for all providers.
-        Only OpenAI and Google providers currently support this method.
-
-        Args:
-            input: Input items (text, files, etc.)
-            pydantic_model: Pydantic model class defining the expected response structure
-            **kwargs: Additional arguments passed to the query method
-
-        Returns:
-            Instance of the pydantic_model with the model's response
-
-        Raises:
-            NotImplementedError: If the provider does not support structured JSON output
-        """
-        raise NotImplementedError(
-            f"query_json is not implemented for {self.__class__.__name__}. "
-            f"Only OpenAI and Google providers currently support this method."
-        )
-
     async def _calculate_cost(
         self,
         metadata: QueryResultMetadata,
@@ -584,7 +321,7 @@ class LLM(ABC):
             )
 
         # costs for long context
-        total_in = metadata.in_tokens + (metadata.cache_read_tokens or 0)
+        total_in = metadata.total_input_tokens
         if costs.context and total_in > costs.context.threshold:
             input_cost, output_cost = costs.context.get_costs(
                 input_cost,
@@ -678,137 +415,29 @@ class LLM(ABC):
         """Upload a file to the model provider"""
         ...
 
-
-class BatchResult(BaseModel):
-    custom_id: str
-    output: QueryResult
-    error_message: str | None = None
-
-
-class LLMBatchMixin(ABC):
-    @abstractmethod
-    async def create_batch_query_request(
+    async def query_json(
         self,
-        custom_id: str,
         input: Sequence[InputItem],
+        pydantic_model: type[PydanticT],
         **kwargs: object,
-    ) -> dict[str, Any]:
-        """Return a single query request
-
-        The batch api sends out a batch of query requests to various endpoints.
+    ) -> PydanticT:
+        """Query the model with JSON response format using Pydantic model.
 
-        For example OpenAI sends can send requests to /v1/responses or /v1/chat/completions endpoints.
+        This is a convenience method that is not implemented for all providers.
+        Only OpenAI and Google providers currently support this method.
 
-        This method creates a query request for methods such methods
-        """
-        ...
+        Args:
+            input: Input items (text, files, etc.)
+            pydantic_model: Pydantic model class defining the expected response structure
+            **kwargs: Additional arguments passed to the query method
 
-    @abstractmethod
-    async def batch_query(
-        self,
-        batch_name: str,
-        requests: list[dict[str, Any]],
-    ) -> str:
-        """
-        Batch query the model
         Returns:
-            str: batch_id
-        Raises:
-            Exception: If failed to batch query
-        """
-        ...
-
-    @abstractmethod
-    async def get_batch_results(self, batch_id: str) -> list[BatchResult]:
-        """
-        Returns results for batch
-        Raises:
-            Exception: If failed to get results
-        """
-        ...
-
-    @abstractmethod
-    async def get_batch_progress(self, batch_id: str) -> int:
-        """
-        Returns number of completed requests for batch
-        Raises:
-            Exception: If failed to get progress
-        """
-        ...
+            Instance of the pydantic_model with the model's response
 
-    @abstractmethod
-    async def cancel_batch_request(self, batch_id: str) -> None:
-        """
-        Cancels batch
         Raises:
-            Exception: If failed to cancel
-        """
-        ...
-
-    @abstractmethod
-    async def get_batch_status(
-        self,
-        batch_id: str,
-    ) -> str:
-        """
-        Returns batch status
-        Raises:
-            Exception: If failed to get status
-        """
-        ...
-
-    @classmethod
-    @abstractmethod
-    def is_batch_status_completed(
-        cls,
-        batch_status: str,
-    ) -> bool:
-        """
-        Returns if batch status is completed
-
-        A completed state is any state that is final and not in-progress
-        Example: failed | cancelled | expired | completed
-
-        An incompleted state is any state that is not completed
-        Example: in_progress | pending | running
+            NotImplementedError: If the provider does not support structured JSON output
         """
-        ...
-
-    @classmethod
-    @abstractmethod
-    def is_batch_status_failed(
-        cls,
-        batch_status: str,
-    ) -> bool:
-        """Returns if batch status is failed"""
-        ...
-
-    @classmethod
-    @abstractmethod
-    def is_batch_status_cancelled(
-        cls,
-        batch_status: str,
-    ) -> bool:
-        """Returns if batch status is cancelled"""
-        ...
-
-
-def get_pretty_input_types(input: Sequence["InputItem"]) -> str:
-    # for logging
-    def process_item(item: "InputItem"):
-        match item:
-            case TextInput():
-                return truncate_str(repr(item))
-            case FileBase():  # FileInput
-                return repr(item)
-            case ToolResult():
-                return repr(item)
-            case dict():
-                item = cast(RawInputItem, item)
-                return repr(item)
-            case _:
-                # RawResponse
-                return repr(item)
-
-    processed_items = [f"  {process_item(item)}" for item in input]
-    return "\n" + "\n".join(processed_items) if processed_items else ""
+        raise NotImplementedError(
+            f"query_json is not implemented for {self.__class__.__name__}. "
+            f"Only OpenAI and Google providers currently support this method."
+        )