vectara-agentic 0.2.12__py3-none-any.whl → 0.2.14__py3-none-any.whl

vectara_agentic/llm_utils.py

@@ -0,0 +1,174 @@
+"""
+Utilities for the Vectara agentic.
+"""
+from types import MethodType
+from typing import Tuple, Callable, Optional
+from functools import lru_cache
+import tiktoken
+
+from llama_index.core.llms import LLM
+from llama_index.llms.openai import OpenAI
+from llama_index.llms.anthropic import Anthropic
+
+from .types import LLMRole, AgentType, ModelProvider
+from .agent_config import AgentConfig
+from .tool_utils import _updated_openai_prepare_chat_with_tools
+
+provider_to_default_model_name = {
+    ModelProvider.OPENAI: "gpt-4o",
+    ModelProvider.ANTHROPIC: "claude-3-7-sonnet-latest",
+    ModelProvider.TOGETHER: "Qwen/Qwen2.5-72B-Instruct-Turbo",
+    ModelProvider.GROQ: "meta-llama/llama-4-scout-17b-16e-instruct",
+    ModelProvider.FIREWORKS: "accounts/fireworks/models/firefunction-v2",
+    ModelProvider.BEDROCK: "anthropic.claude-3-7-sonnet-20250219-v1:0",
+    ModelProvider.COHERE: "command-a-03-2025",
+    ModelProvider.GEMINI: "models/gemini-2.0-flash",
+}
+
+DEFAULT_MODEL_PROVIDER = ModelProvider.OPENAI
+
+
+@lru_cache(maxsize=None)
+def _get_llm_params_for_role(
+    role: LLMRole, config: Optional[AgentConfig] = None
+) -> Tuple[ModelProvider, str]:
+    """
+    Get the model provider and model name for the specified role.
+
+    If config is None, a new AgentConfig() is instantiated using environment defaults.
+    """
+    config = config or AgentConfig()  # fallback to default config
+
+    if role == LLMRole.TOOL:
+        model_provider = ModelProvider(config.tool_llm_provider)
+        # If the user hasn’t explicitly set a tool_llm_model_name,
+        # fallback to provider default from provider_to_default_model_name
+        model_name = config.tool_llm_model_name or provider_to_default_model_name.get(
+            model_provider
+        )
+    else:
+        model_provider = ModelProvider(config.main_llm_provider)
+        model_name = config.main_llm_model_name or provider_to_default_model_name.get(
+            model_provider
+        )
+
+    # If the agent type is OpenAI, check that the main LLM provider is also OpenAI.
+    if role == LLMRole.MAIN and config.agent_type == AgentType.OPENAI:
+        if model_provider != ModelProvider.OPENAI:
+            raise ValueError(
+                "OpenAI agent requested but main model provider is not OpenAI."
+            )
+
+    return model_provider, model_name
+
+
+@lru_cache(maxsize=None)
+def get_tokenizer_for_model(
+    role: LLMRole, config: Optional[AgentConfig] = None
+) -> Optional[Callable]:
+    """
+    Get the tokenizer for the specified model, as determined by the role & config.
+    """
+    model_provider, model_name = _get_llm_params_for_role(role, config)
+    if model_provider == ModelProvider.OPENAI:
+        # This might raise an exception if the model_name is unknown to tiktoken
+        return tiktoken.encoding_for_model(model_name).encode
+    if model_provider == ModelProvider.ANTHROPIC:
+        return Anthropic().tokenizer
+    return None
+
+
+@lru_cache(maxsize=None)
+def get_llm(role: LLMRole, config: Optional[AgentConfig] = None) -> LLM:
+    """
+    Get the LLM for the specified role, using the provided config
+    or a default if none is provided.
+    """
+    max_tokens = 8192
+    model_provider, model_name = _get_llm_params_for_role(role, config)
+    if model_provider == ModelProvider.OPENAI:
+        llm = OpenAI(
+            model=model_name,
+            temperature=0,
+            is_function_calling_model=True,
+            strict=True,
+            max_tokens=max_tokens,
+            pydantic_program_mode="openai",
+        )
+    elif model_provider == ModelProvider.ANTHROPIC:
+        llm = Anthropic(
+            model=model_name,
+            temperature=0,
+            max_tokens=max_tokens,
+        )
+    elif model_provider == ModelProvider.GEMINI:
+        from llama_index.llms.google_genai import GoogleGenAI
+
+        llm = GoogleGenAI(
+            model=model_name,
+            temperature=0,
+            is_function_calling_model=True,
+            allow_parallel_tool_calls=True,
+            max_tokens=max_tokens,
+        )
+    elif model_provider == ModelProvider.TOGETHER:
+        from llama_index.llms.together import TogetherLLM
+
+        llm = TogetherLLM(
+            model=model_name,
+            temperature=0,
+            is_function_calling_model=True,
+            max_tokens=max_tokens,
+        )
+        # pylint: disable=protected-access
+        llm._prepare_chat_with_tools = MethodType(
+            _updated_openai_prepare_chat_with_tools,
+            llm,
+        )
+    elif model_provider == ModelProvider.GROQ:
+        from llama_index.llms.groq import Groq
+
+        llm = Groq(
+            model=model_name,
+            temperature=0,
+            is_function_calling_model=True,
+            max_tokens=max_tokens,
+        )
+        # pylint: disable=protected-access
+        llm._prepare_chat_with_tools = MethodType(
+            _updated_openai_prepare_chat_with_tools,
+            llm,
+        )
+    elif model_provider == ModelProvider.FIREWORKS:
+        from llama_index.llms.fireworks import Fireworks
+
+        llm = Fireworks(model=model_name, temperature=0, max_tokens=max_tokens)
+    elif model_provider == ModelProvider.BEDROCK:
+        from llama_index.llms.bedrock import Bedrock
+
+        llm = Bedrock(model=model_name, temperature=0, max_tokens=max_tokens)
+    elif model_provider == ModelProvider.COHERE:
+        from llama_index.llms.cohere import Cohere
+
+        llm = Cohere(model=model_name, temperature=0, max_tokens=max_tokens)
+    elif model_provider == ModelProvider.PRIVATE:
+        from llama_index.llms.openai_like import OpenAILike
+
+        llm = OpenAILike(
+            model=model_name,
+            temperature=0,
+            is_function_calling_model=True,
+            is_chat_model=True,
+            api_base=config.private_llm_api_base,
+            api_key=config.private_llm_api_key,
+            max_tokens=max_tokens,
+        )
+        # pylint: disable=protected-access
+        llm._prepare_chat_with_tools = MethodType(
+            _updated_openai_prepare_chat_with_tools,
+            llm,
+        )
+
+    else:
+        raise ValueError(f"Unknown LLM provider: {model_provider}")
+    return llm

vectara_agentic/tool_utils.py

@@ -0,0 +1,513 @@
+"""
+This module contains the ToolsFactory class for creating agent tools.
+"""
+
+import inspect
+import re
+
+from typing import (
+    Callable, List, Dict, Any, Optional, Union, Type, Tuple,
+    Sequence
+)
+from pydantic import BaseModel, create_model
+from pydantic_core import PydanticUndefined
+
+from llama_index.core.tools import FunctionTool
+from llama_index.core.tools.function_tool import AsyncCallable
+from llama_index.core.tools.types import ToolMetadata, ToolOutput
+from llama_index.core.workflow.context import Context
+
+from llama_index.core.tools.types import BaseTool
+from llama_index.core.base.llms.types import ChatMessage, MessageRole
+from llama_index.llms.openai.utils import resolve_tool_choice
+
+from .types import ToolType
+from .utils import is_float
+
+
+def _updated_openai_prepare_chat_with_tools(
+    self,
+    tools: Sequence["BaseTool"],
+    user_msg: Optional[Union[str, ChatMessage]] = None,
+    chat_history: Optional[List[ChatMessage]] = None,
+    verbose: bool = False,
+    allow_parallel_tool_calls: bool = False,
+    tool_choice: Union[str, dict] = "auto",
+    strict: Optional[bool] = None,
+    **kwargs: Any,
+) -> Dict[str, Any]:
+    """Predict and call the tool."""
+    tool_specs = [tool.metadata.to_openai_tool(skip_length_check=True) for tool in tools]
+
+    # if strict is passed in, use, else default to the class-level attribute, else default to True`
+    strict = strict if strict is not None else self.strict
+
+    if self.metadata.is_function_calling_model:
+        for tool_spec in tool_specs:
+            if tool_spec["type"] == "function":
+                tool_spec["function"]["strict"] = strict
+                # in current openai 1.40.0 it is always false.
+                tool_spec["function"]["parameters"]["additionalProperties"] = False
+
+    if isinstance(user_msg, str):
+        user_msg = ChatMessage(role=MessageRole.USER, content=user_msg)
+
+    messages = chat_history or []
+    if user_msg:
+        messages.append(user_msg)
+
+    return {
+        "messages": messages,
+        "tools": tool_specs or None,
+        "tool_choice": resolve_tool_choice(tool_choice) if tool_specs else None,
+        **kwargs,
+    }
+
+class VectaraToolMetadata(ToolMetadata):
+    """
+    A subclass of ToolMetadata adding the tool_type attribute.
+    """
+
+    tool_type: ToolType
+
+    def __init__(self, tool_type: ToolType, **kwargs):
+        super().__init__(**kwargs)
+        self.tool_type = tool_type
+
+    def __repr__(self) -> str:
+        """
+        Returns a string representation of the VectaraToolMetadata object, including the tool_type attribute.
+        """
+        base_repr = super().__repr__()
+        return f"{base_repr}, tool_type={self.tool_type}"
+
+
+class VectaraTool(FunctionTool):
+    """
+    A subclass of FunctionTool adding the tool_type attribute.
+    """
+
+    def __init__(
+        self,
+        tool_type: ToolType,
+        metadata: ToolMetadata,
+        fn: Optional[Callable[..., Any]] = None,
+        async_fn: Optional[AsyncCallable] = None,
+    ) -> None:
+        metadata_dict = (
+            metadata.dict() if hasattr(metadata, "dict") else metadata.__dict__
+        )
+        vm = VectaraToolMetadata(tool_type=tool_type, **metadata_dict)
+        super().__init__(fn, vm, async_fn)
+
+    @classmethod
+    def from_defaults(
+        cls,
+        fn: Optional[Callable[..., Any]] = None,
+        name: Optional[str] = None,
+        description: Optional[str] = None,
+        return_direct: bool = False,
+        fn_schema: Optional[Type[BaseModel]] = None,
+        async_fn: Optional[AsyncCallable] = None,
+        tool_metadata: Optional[ToolMetadata] = None,
+        callback: Optional[Callable[[Any], Any]] = None,
+        async_callback: Optional[AsyncCallable] = None,
+        tool_type: ToolType = ToolType.QUERY,
+    ) -> "VectaraTool":
+        tool = FunctionTool.from_defaults(
+            fn,
+            name,
+            description,
+            return_direct,
+            fn_schema,
+            async_fn,
+            tool_metadata,
+            callback,
+            async_callback,
+        )
+        vectara_tool = cls(
+            tool_type=tool_type,
+            fn=tool.fn,
+            metadata=tool.metadata,
+            async_fn=tool.async_fn,
+        )
+        return vectara_tool
+
+    def __str__(self) -> str:
+        return f"Tool(name={self.metadata.name}, " f"Tool metadata={self.metadata})"
+
+    def __repr__(self) -> str:
+        return str(self)
+
+    def __eq__(self, other):
+        if not isinstance(other, VectaraTool):
+            return False
+
+        if self.metadata.tool_type != other.metadata.tool_type:
+            return False
+
+        if self.metadata.name != other.metadata.name:
+            return False
+
+        # If schema is a dict-like object, compare the dict representation
+        try:
+            # Try to get schema as dict if possible
+            if hasattr(self.metadata.fn_schema, "schema"):
+                self_schema = self.metadata.fn_schema.schema
+                other_schema = other.metadata.fn_schema.schema
+
+                # Compare only properties and required fields
+                self_props = self_schema.get("properties", {})
+                other_props = other_schema.get("properties", {})
+
+                self_required = self_schema.get("required", [])
+                other_required = other_schema.get("required", [])
+
+                return self_props.keys() == other_props.keys() and set(
+                    self_required
+                ) == set(other_required)
+        except Exception:
+            # If any exception occurs during schema comparison, fall back to name comparison
+            pass
+
+        return True
+
+    def call(
+        self, *args: Any, ctx: Optional[Context] = None, **kwargs: Any
+    ) -> ToolOutput:
+        try:
+            return super().call(*args, ctx=ctx, **kwargs)
+        except TypeError as e:
+            sig = inspect.signature(self.metadata.fn_schema)
+            valid_parameters = list(sig.parameters.keys())
+            params_str = ", ".join(valid_parameters)
+
+            err_output = ToolOutput(
+                tool_name=self.metadata.name,
+                content=(
+                    f"Wrong argument used when calling {self.metadata.name}: {str(e)}. "
+                    f"Valid arguments: {params_str}. please call the tool again with the correct arguments."
+                ),
+                raw_input={"args": args, "kwargs": kwargs},
+                raw_output={"response": str(e)},
+            )
+            return err_output
+        except Exception as e:
+            err_output = ToolOutput(
+                tool_name=self.metadata.name,
+                content=f"Tool {self.metadata.name} Malfunction: {str(e)}",
+                raw_input={"args": args, "kwargs": kwargs},
+                raw_output={"response": str(e)},
+            )
+            return err_output
+
+    async def acall(
+        self, *args: Any, ctx: Optional[Context] = None, **kwargs: Any
+    ) -> ToolOutput:
+        try:
+            return await super().acall(*args, ctx=ctx, **kwargs)
+        except TypeError as e:
+            sig = inspect.signature(self.metadata.fn_schema)
+            valid_parameters = list(sig.parameters.keys())
+            params_str = ", ".join(valid_parameters)
+
+            err_output = ToolOutput(
+                tool_name=self.metadata.name,
+                content=(
+                    f"Wrong argument used when calling {self.metadata.name}: {str(e)}. "
+                    f"Valid arguments: {params_str}. please call the tool again with the correct arguments."
+                ),
+                raw_input={"args": args, "kwargs": kwargs},
+                raw_output={"response": str(e)},
+            )
+            return err_output
+        except Exception as e:
+            err_output = ToolOutput(
+                tool_name=self.metadata.name,
+                content=f"Tool {self.metadata.name} Malfunction: {str(e)}",
+                raw_input={"args": args, "kwargs": kwargs},
+                raw_output={"response": str(e)},
+            )
+            return err_output
+
+
+def _create_tool_from_dynamic_function(
+    function: Callable[..., ToolOutput],
+    tool_name: str,
+    tool_description: str,
+    base_params_model: Type[BaseModel],
+    tool_args_schema: Type[BaseModel],
+    compact_docstring: bool = False,
+) -> VectaraTool:
+    base_params = []
+
+    if tool_args_schema is None:
+
+        class EmptyBaseModel(BaseModel):
+            """empty base model"""
+
+        tool_args_schema = EmptyBaseModel
+
+    fields = {}
+    for param_name, model_field in base_params_model.model_fields.items():
+        field_type = base_params_model.__annotations__.get(
+            param_name, str
+        )
+        default_value = (
+            model_field.default
+            if model_field.default is not None
+            else inspect.Parameter.empty
+        )
+        param = inspect.Parameter(
+            param_name,
+            inspect.Parameter.POSITIONAL_OR_KEYWORD,
+            default=default_value,
+            annotation=field_type,
+        )
+        base_params.append(param)
+        fields[param_name] = (
+            field_type,
+            model_field.default if model_field.default is not None else ...,
+        )
+
+    # Add tool_args_schema fields to the fields dict if not already included.
+    # Also add them to the function signature by creating new inspect.Parameter objects.
+    for field_name, field_info in tool_args_schema.model_fields.items():
+        if field_name in fields:
+            continue
+
+        default_value = field_info.default if field_info.default is not None else ...
+        field_type = tool_args_schema.__annotations__.get(field_name, None)
+        fields[field_name] = (field_type, default_value)
+        param = inspect.Parameter(
+            field_name,
+            inspect.Parameter.POSITIONAL_OR_KEYWORD,
+            default=(
+                default_value
+                if default_value is not ...
+                else inspect.Parameter.empty
+            ),
+            annotation=field_type,
+        )
+        base_params.append(param)
+
+    # Create the dynamic schema with both base_params_model and tool_args_schema fields.
+    fn_schema = create_model(f"{tool_name}_schema", **fields)
+
+    # Combine parameters into a function signature.
+    all_params = base_params[:]  # Now all_params contains parameters from both models.
+    required_params = [p for p in all_params if p.default is inspect.Parameter.empty]
+    optional_params = [
+        p for p in all_params if p.default is not inspect.Parameter.empty
+    ]
+    function.__signature__ = inspect.Signature(required_params + optional_params)
+    function.__annotations__["return"] = dict[str, Any]
+    function.__name__ = re.sub(r"[^A-Za-z0-9_]", "_", tool_name)
+
+    # Build a docstring using parameter descriptions from the BaseModels.
+    params_str = ", ".join(
+        f"{p.name}: {p.annotation.__name__ if hasattr(p.annotation, '__name__') else p.annotation}"
+        for p in all_params
+    )
+    signature_line = f"{tool_name}({params_str}) -> dict[str, Any]"
+    if compact_docstring:
+        doc_lines = [
+            tool_description.strip(),
+        ]
+    else:
+        doc_lines = [
+            signature_line,
+            "",
+            tool_description.strip(),
+        ]
+    doc_lines += [
+        "",
+        "Args:",
+    ]
+    for param in all_params:
+        description = ""
+        if param.name in base_params_model.model_fields:
+            description = base_params_model.model_fields[param.name].description
+        elif param.name in tool_args_schema.model_fields:
+            description = tool_args_schema.model_fields[param.name].description
+        if not description:
+            description = ""
+        type_name = (
+            param.annotation.__name__
+            if hasattr(param.annotation, "__name__")
+            else str(param.annotation)
+        )
+        if (
+            param.default is not inspect.Parameter.empty
+            and param.default is not PydanticUndefined
+        ):
+            default_text = f", default={param.default!r}"
+        else:
+            default_text = ""
+        doc_lines.append(f"  - {param.name} ({type_name}){default_text}: {description}")
+    doc_lines.append("")
+    doc_lines.append("Returns:")
+    return_desc = getattr(
+        function, "__return_description__", "A dictionary containing the result data."
+    )
+    doc_lines.append(f"    dict[str, Any]: {return_desc}")
+
+    initial_docstring = "\n".join(doc_lines)
+    collapsed_spaces = re.sub(r' {2,}', ' ', initial_docstring)
+    final_docstring = re.sub(r'\n{2,}', '\n', collapsed_spaces).strip()
+    function.__doc__ = final_docstring
+
+    tool = VectaraTool.from_defaults(
+        fn=function,
+        name=tool_name,
+        description=function.__doc__,
+        fn_schema=fn_schema,
+        tool_type=ToolType.QUERY,
+    )
+    return tool
+
+
+Range = Tuple[float, float, bool, bool]  # (min, max, min_inclusive, max_inclusive)
+
+
+def _parse_range(val_str: str) -> Range:
+    """
+    Parses '[1,10)' or '(0.5, 5]' etc.
+    Returns (start, end, start_incl, end_incl) or raises ValueError.
+    """
+    m = re.match(
+        r"""
+        ^([\[\(])\s*            # opening bracket
+        ([+-]?\d+(\.\d*)?)\s*,  # first number
+        \s*([+-]?\d+(\.\d*)?)   # second number
+        \s*([\]\)])$            # closing bracket
+    """,
+        val_str,
+        re.VERBOSE,
+    )
+    if not m:
+        raise ValueError(f"Invalid range syntax: {val_str!r}")
+    start_inc = m.group(1) == "["
+    end_inc = m.group(7) == "]"
+    start = float(m.group(2))
+    end = float(m.group(4))
+    if start > end:
+        raise ValueError(f"Range lower bound greater than upper bound: {val_str!r}")
+    return start, end, start_inc, end_inc
+
+
+def _parse_comparison(val_str: str) -> Tuple[str, Union[float, str, bool]]:
+    """
+    Parses '>10', '<=3.14', '!=foo', \"='bar'\" etc.
+    Returns (operator, rhs) or raises ValueError.
+    """
+    # pick off the operator
+    comparison_operators = [">=", "<=", "!=", ">", "<", "="]
+    numeric_only_operators = {">", "<", ">=", "<="}
+    for op in comparison_operators:
+        if val_str.startswith(op):
+            rhs = val_str[len(op) :].strip()
+            if op in numeric_only_operators:
+                try:
+                    rhs_val = float(rhs)
+                except ValueError as e:
+                    raise ValueError(
+                        f"Numeric comparison {op!r} must have a number, got {rhs!r}"
+                    ) from e
+                return op, rhs_val
+            # = and != can be bool, numeric, or string
+            low = rhs.lower()
+            if low in ("true", "false"):
+                return op, (low == "true")
+            try:
+                return op, float(rhs)
+            except ValueError:
+                return op, rhs
+    raise ValueError(f"No valid comparison operator at start of {val_str!r}")
+
+
+def _build_filter_string(
+    kwargs: Dict[str, Any], tool_args_type: Dict[str, dict], fixed_filter: str
+) -> str:
+    """
+    Build filter string for Vectara from kwargs
+    """
+    filter_parts = []
+    for key, raw in kwargs.items():
+        if raw is None or raw == "":
+            continue
+
+        if raw is PydanticUndefined:
+            raise ValueError(
+                f"Value of argument {key!r} is undefined, and this is invalid. "
+                "Please form proper arguments and try again."
+            )
+
+        tool_args_dict = tool_args_type.get(key, {"type": "doc", "is_list": False})
+        prefix = tool_args_dict.get("type", "doc")
+        is_list = tool_args_dict.get("is_list", False)
+
+        if prefix not in ("doc", "part"):
+            raise ValueError(
+                f'Unrecognized prefix {prefix!r}. Please make sure to use either "doc" or "part" for the prefix.'
+            )
+
+        # 1) native numeric
+        if isinstance(raw, (int, float)) or is_float(str(raw)):
+            val = str(raw)
+            if is_list:
+                filter_parts.append(f"({val} IN {prefix}.{key})")
+            else:
+                filter_parts.append(f"{prefix}.{key}={val}")
+            continue
+
+        # 2) native boolean
+        if isinstance(raw, bool):
+            val = "true" if raw else "false"
+            if is_list:
+                filter_parts.append(f"({val} IN {prefix}.{key})")
+            else:
+                filter_parts.append(f"{prefix}.{key}={val}")
+            continue
+
+        if not isinstance(raw, str):
+            raise ValueError(f"Unsupported type for {key!r}: {type(raw).__name__}")
+
+        val_str = raw.strip()
+
+        # 3) Range operator
+        if (val_str.startswith("[") or val_str.startswith("(")) and (
+            val_str.endswith("]") or val_str.endswith(")")
+        ):
+            start, end, start_incl, end_incl = _parse_range(val_str)
+            conds = []
+            op1 = ">=" if start_incl else ">"
+            op2 = "<=" if end_incl else "<"
+            conds.append(f"{prefix}.{key} {op1} {start}")
+            conds.append(f"{prefix}.{key} {op2} {end}")
+            filter_parts.append("(" + " AND ".join(conds) + ")")
+            continue
+
+        # 4) comparison operator
+        try:
+            op, rhs = _parse_comparison(val_str)
+        except ValueError:
+            # no operator → treat as membership or equality-on-string
+            if is_list:
+                filter_parts.append(f"('{val_str}' IN {prefix}.{key})")
+            else:
+                filter_parts.append(f"{prefix}.{key}='{val_str}'")
+        else:
+            # normal comparison always binds to the field
+            if isinstance(rhs, bool):
+                rhs_sql = "true" if rhs else "false"
+            elif isinstance(rhs, (int, float)):
+                rhs_sql = str(rhs)
+            else:
+                rhs_sql = f"'{rhs}'"
+            filter_parts.append(f"{prefix}.{key}{op}{rhs_sql}")
+
+    joined = " AND ".join(filter_parts)
+    if fixed_filter and joined:
+        return f"({fixed_filter}) AND ({joined})"
+    return fixed_filter or joined