community: support Databricks Unity Catalog functions as LangChain tools (#22555)

This PR adds support for using Databricks Unity Catalog functions as
LangChain tools, which runs inside a Databricks SQL warehouse.

* An example notebook is provided.

docs/docs/integrations/tools/databricks.ipynb

@@ -0,0 +1,168 @@
+{
+ "cells": [
+  {
+   "cell_type": "markdown",
+   "metadata": {},
+   "source": [
+    "# Databricks Unity Catalog (UC)\n",
+    "\n",
+    "This notebook shows how to use UC functions as LangChain tools.\n",
+    "\n",
+    "See Databricks documentation ([AWS](https://docs.databricks.com/en/sql/language-manual/sql-ref-syntax-ddl-create-sql-function.html)|[Azure](https://learn.microsoft.com/en-us/azure/databricks/sql/language-manual/sql-ref-syntax-ddl-create-sql-function)|[GCP](https://docs.gcp.databricks.com/en/sql/language-manual/sql-ref-syntax-ddl-create-sql-function.html)) to learn how to create SQL or Python functions in UC. Do not skip function and parameter comments, which are critical for LLMs to call functions properly.\n",
+    "\n",
+    "In this example notebook, we create a simple Python function that executes arbitary code and use it as a LangChain tool:\n",
+    "\n",
+    "```sql\n",
+    "CREATE FUNCTION main.tools.python_exec (\n",
+    "  code STRING COMMENT 'Python code to execute. Remember to print the final result to stdout.'\n",
+    ")\n",
+    "RETURNS STRING\n",
+    "LANGUAGE PYTHON\n",
+    "COMMENT 'Executes Python code and returns its stdout.'\n",
+    "AS $$\n",
+    "  import sys\n",
+    "  from io import StringIO\n",
+    "  stdout = StringIO()\n",
+    "  sys.stdout = stdout\n",
+    "  exec(code)\n",
+    "  return stdout.getvalue()\n",
+    "$$\n",
+    "```\n",
+    "\n",
+    "It runs in a secure and isolated environment within a Databricks SQL warehouse."
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "%pip install --upgrade --quiet databricks-sdk langchain-community langchain-openai"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 2,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_openai import ChatOpenAI\n",
+    "\n",
+    "llm = ChatOpenAI(model=\"gpt-3.5-turbo\")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 3,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain_community.tools.databricks import UCFunctionToolkit\n",
+    "\n",
+    "tools = (\n",
+    "    UCFunctionToolkit(\n",
+    "        # You can find the SQL warehouse ID in its UI after creation.\n",
+    "        warehouse_id=\"xxxx123456789\"\n",
+    "    )\n",
+    "    .include(\n",
+    "        # Include functions as tools using their qualified names.\n",
+    "        # You can use \"{catalog_name}.{schema_name}.*\" to get all functions in a schema.\n",
+    "        \"main.tools.python_exec\",\n",
+    "    )\n",
+    "    .get_tools()\n",
+    ")"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 4,
+   "metadata": {},
+   "outputs": [],
+   "source": [
+    "from langchain.agents import AgentExecutor, create_tool_calling_agent\n",
+    "from langchain_core.prompts import ChatPromptTemplate\n",
+    "\n",
+    "prompt = ChatPromptTemplate.from_messages(\n",
+    "    [\n",
+    "        (\n",
+    "            \"system\",\n",
+    "            \"You are a helpful assistant. Make sure to use tool for information.\",\n",
+    "        ),\n",
+    "        (\"placeholder\", \"{chat_history}\"),\n",
+    "        (\"human\", \"{input}\"),\n",
+    "        (\"placeholder\", \"{agent_scratchpad}\"),\n",
+    "    ]\n",
+    ")\n",
+    "\n",
+    "agent = create_tool_calling_agent(llm, tools, prompt)"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": 5,
+   "metadata": {},
+   "outputs": [
+    {
+     "name": "stdout",
+     "output_type": "stream",
+     "text": [
+      "\n",
+      "\n",
+      "\u001b[1m> Entering new AgentExecutor chain...\u001b[0m\n",
+      "\u001b[32;1m\u001b[1;3m\n",
+      "Invoking: `main__tools__python_exec` with `{'code': 'print(36939 * 8922.4)'}`\n",
+      "\n",
+      "\n",
+      "\u001b[0m\u001b[36;1m\u001b[1;3m{\"format\": \"SCALAR\", \"value\": \"329584533.59999996\\n\", \"truncated\": false}\u001b[0m\u001b[32;1m\u001b[1;3mThe result of the multiplication 36939 * 8922.4 is 329,584,533.60.\u001b[0m\n",
+      "\n",
+      "\u001b[1m> Finished chain.\u001b[0m\n"
+     ]
+    },
+    {
+     "data": {
+      "text/plain": [
+       "{'input': '36939 * 8922.4',\n",
+       " 'output': 'The result of the multiplication 36939 * 8922.4 is 329,584,533.60.'}"
+      ]
+     },
+     "execution_count": 5,
+     "metadata": {},
+     "output_type": "execute_result"
+    }
+   ],
+   "source": [
+    "agent_executor = AgentExecutor(agent=agent, tools=tools, verbose=True)\n",
+    "agent_executor.invoke({\"input\": \"36939 * 8922.4\"})"
+   ]
+  },
+  {
+   "cell_type": "code",
+   "execution_count": null,
+   "metadata": {},
+   "outputs": [],
+   "source": []
+  }
+ ],
+ "metadata": {
+  "kernelspec": {
+   "display_name": "llm",
+   "language": "python",
+   "name": "python3"
+  },
+  "language_info": {
+   "codemirror_mode": {
+    "name": "ipython",
+    "version": 3
+   },
+   "file_extension": ".py",
+   "mimetype": "text/x-python",
+   "name": "python",
+   "nbconvert_exporter": "python",
+   "pygments_lexer": "ipython3",
+   "version": "3.11.9"
+  }
+ },
+ "nbformat": 4,
+ "nbformat_minor": 2
+}

libs/community/langchain_community/tools/databricks/__init__.py

@@ -0,0 +1,3 @@
+from langchain_community.tools.databricks.tool import UCFunctionToolkit
+
+__all__ = ["UCFunctionToolkit"]

libs/community/langchain_community/tools/databricks/_execution.py

@@ -0,0 +1,172 @@
+import json
+from dataclasses import dataclass
+from io import StringIO
+from typing import TYPE_CHECKING, Any, Dict, List, Literal, Optional
+
+if TYPE_CHECKING:
+    from databricks.sdk import WorkspaceClient
+    from databricks.sdk.service.catalog import FunctionInfo
+    from databricks.sdk.service.sql import StatementParameterListItem
+
+
+def is_scalar(function: "FunctionInfo") -> bool:
+    from databricks.sdk.service.catalog import ColumnTypeName
+
+    return function.data_type != ColumnTypeName.TABLE_TYPE
+
+
+@dataclass
+class ParameterizedStatement:
+    statement: str
+    parameters: List["StatementParameterListItem"]
+
+
+@dataclass
+class FunctionExecutionResult:
+    """
+    Result of executing a function.
+    We always use a string to present the result value for AI model to consume.
+    """
+
+    error: Optional[str] = None
+    format: Optional[Literal["SCALAR", "CSV"]] = None
+    value: Optional[str] = None
+    truncated: Optional[bool] = None
+
+    def to_json(self) -> str:
+        data = {k: v for (k, v) in self.__dict__.items() if v is not None}
+        return json.dumps(data)
+
+
+def get_execute_function_sql_stmt(
+    function: "FunctionInfo", json_params: Dict[str, Any]
+) -> ParameterizedStatement:
+    from databricks.sdk.service.catalog import ColumnTypeName
+    from databricks.sdk.service.sql import StatementParameterListItem
+
+    parts = []
+    output_params = []
+    if is_scalar(function):
+        # TODO: IDENTIFIER(:function) did not work
+        parts.append(f"SELECT {function.full_name}(")
+    else:
+        parts.append(f"SELECT * FROM {function.full_name}(")
+    if function.input_params is None or function.input_params.parameters is None:
+        assert (
+            not json_params
+        ), "Function has no parameters but parameters were provided."
+    else:
+        args = []
+        use_named_args = False
+        for p in function.input_params.parameters:
+            if p.name not in json_params:
+                if p.parameter_default is not None:
+                    use_named_args = True
+                else:
+                    raise ValueError(
+                        f"Parameter {p.name} is required but not provided."
+                    )
+            else:
+                arg_clause = ""
+                if use_named_args:
+                    arg_clause += f"{p.name} => "
+                json_value = json_params[p.name]
+                if p.type_name in (
+                    ColumnTypeName.ARRAY,
+                    ColumnTypeName.MAP,
+                    ColumnTypeName.STRUCT,
+                ):
+                    # Use from_json to restore values of complex types.
+                    json_value_str = json.dumps(json_value)
+                    # TODO: parametrize type
+                    arg_clause += f"from_json(:{p.name}, '{p.type_text}')"
+                    output_params.append(
+                        StatementParameterListItem(name=p.name, value=json_value_str)
+                    )
+                elif p.type_name == ColumnTypeName.BINARY:
+                    # Use ubbase64 to restore binary values.
+                    arg_clause += f"unbase64(:{p.name})"
+                    output_params.append(
+                        StatementParameterListItem(name=p.name, value=json_value)
+                    )
+                else:
+                    arg_clause += f":{p.name}"
+                    output_params.append(
+                        StatementParameterListItem(
+                            name=p.name, value=json_value, type=p.type_text
+                        )
+                    )
+                args.append(arg_clause)
+        parts.append(",".join(args))
+    parts.append(")")
+    # TODO: check extra params in kwargs
+    statement = "".join(parts)
+    return ParameterizedStatement(statement=statement, parameters=output_params)
+
+
+def execute_function(
+    ws: "WorkspaceClient",
+    warehouse_id: str,
+    function: "FunctionInfo",
+    parameters: Dict[str, Any],
+) -> FunctionExecutionResult:
+    """
+    Execute a function with the given arguments and return the result.
+    """
+    try:
+        import pandas as pd
+    except ImportError as e:
+        raise ImportError(
+            "Could not import pandas python package. "
+            "Please install it with `pip install pandas`."
+        ) from e
+    from databricks.sdk.service.sql import StatementState
+
+    # TODO: async so we can run functions in parallel
+    parametrized_statement = get_execute_function_sql_stmt(function, parameters)
+    # TODO: configurable limits
+    response = ws.statement_execution.execute_statement(
+        statement=parametrized_statement.statement,
+        warehouse_id=warehouse_id,
+        parameters=parametrized_statement.parameters,
+        wait_timeout="30s",
+        row_limit=100,
+        byte_limit=4096,
+    )
+    status = response.status
+    assert status is not None, f"Statement execution failed: {response}"
+    if status.state != StatementState.SUCCEEDED:
+        error = status.error
+        assert (
+            error is not None
+        ), "Statement execution failed but no error message was provided."
+        return FunctionExecutionResult(error=f"{error.error_code}: {error.message}")
+    manifest = response.manifest
+    assert manifest is not None
+    truncated = manifest.truncated
+    result = response.result
+    assert (
+        result is not None
+    ), "Statement execution succeeded but no result was provided."
+    data_array = result.data_array
+    if is_scalar(function):
+        value = None
+        if data_array and len(data_array) > 0 and len(data_array[0]) > 0:
+            value = str(data_array[0][0])  # type: ignore
+        return FunctionExecutionResult(
+            format="SCALAR", value=value, truncated=truncated
+        )
+    else:
+        schema = manifest.schema
+        assert (
+            schema is not None and schema.columns is not None
+        ), "Statement execution succeeded but no schema was provided."
+        columns = [c.name for c in schema.columns]
+        if data_array is None:
+            data_array = []
+        pdf = pd.DataFrame.from_records(data_array, columns=columns)
+        csv_buffer = StringIO()
+        pdf.to_csv(csv_buffer, index=False)
+        return FunctionExecutionResult(
+            format="CSV", value=csv_buffer.getvalue(), truncated=truncated
+        )

libs/community/langchain_community/tools/databricks/tool.py

@@ -0,0 +1,201 @@
+import json
+from datetime import date, datetime
+from decimal import Decimal
+from hashlib import md5
+from typing import TYPE_CHECKING, Any, Dict, List, Optional, Type, Union
+
+from langchain_core.pydantic_v1 import BaseModel, Field, create_model
+from langchain_core.tools import BaseTool, BaseToolkit, StructuredTool
+from typing_extensions import Self
+
+if TYPE_CHECKING:
+    from databricks.sdk import WorkspaceClient
+    from databricks.sdk.service.catalog import FunctionInfo
+
+from langchain_community.tools.databricks._execution import execute_function
+
+
+def _uc_type_to_pydantic_type(uc_type_json: Union[str, Dict[str, Any]]) -> Type:
+    mapping = {
+        "long": int,
+        "binary": bytes,
+        "boolean": bool,
+        "date": date,
+        "double": float,
+        "float": float,
+        "integer": int,
+        "short": int,
+        "string": str,
+        "timestamp": datetime,
+        "timestamp_ntz": datetime,
+        "byte": int,
+    }
+    if isinstance(uc_type_json, str):
+        if uc_type_json in mapping:
+            return mapping[uc_type_json]
+        else:
+            if uc_type_json.startswith("decimal"):
+                return Decimal
+            elif uc_type_json == "void" or uc_type_json.startswith("interval"):
+                raise TypeError(f"Type {uc_type_json} is not supported.")
+            else:
+                raise TypeError(
+                    f"Unknown type {uc_type_json}. Try upgrading this package."
+                )
+    else:
+        assert isinstance(uc_type_json, dict)
+        tpe = uc_type_json["type"]
+        if tpe == "array":
+            element_type = _uc_type_to_pydantic_type(uc_type_json["elementType"])
+            if uc_type_json["containsNull"]:
+                element_type = Optional[element_type]  # type: ignore
+            return List[element_type]  # type: ignore
+        elif tpe == "map":
+            key_type = uc_type_json["keyType"]
+            assert key_type == "string", TypeError(
+                f"Only support STRING key type for MAP but got {key_type}."
+            )
+            value_type = _uc_type_to_pydantic_type(uc_type_json["valueType"])
+            if uc_type_json["valueContainsNull"]:
+                value_type: Type = Optional[value_type]  # type: ignore
+            return Dict[str, value_type]  # type: ignore
+        elif tpe == "struct":
+            fields = {}
+            for field in uc_type_json["fields"]:
+                field_type = _uc_type_to_pydantic_type(field["type"])
+                if field.get("nullable"):
+                    field_type = Optional[field_type]  # type: ignore
+                comment = (
+                    uc_type_json["metadata"].get("comment")
+                    if "metadata" in uc_type_json
+                    else None
+                )
+                fields[field["name"]] = (field_type, Field(..., description=comment))
+            uc_type_json_str = json.dumps(uc_type_json, sort_keys=True)
+            type_hash = md5(uc_type_json_str.encode()).hexdigest()[:8]
+            return create_model(f"Struct_{type_hash}", **fields)  # type: ignore
+        else:
+            raise TypeError(f"Unknown type {uc_type_json}. Try upgrading this package.")
+
+
+def _generate_args_schema(function: "FunctionInfo") -> Type[BaseModel]:
+    if function.input_params is None:
+        return BaseModel
+    params = function.input_params.parameters
+    assert params is not None
+    fields = {}
+    for p in params:
+        assert p.type_json is not None
+        type_json = json.loads(p.type_json)["type"]
+        pydantic_type = _uc_type_to_pydantic_type(type_json)
+        description = p.comment
+        default: Any = ...
+        if p.parameter_default:
+            pydantic_type = Optional[pydantic_type]  # type: ignore
+            default = None
+            # TODO: Convert default value string to the correct type.
+            # We might need to use statement execution API
+            # to get the JSON representation of the value.
+            default_description = f"(Default: {p.parameter_default})"
+            if description:
+                description += f" {default_description}"
+            else:
+                description = default_description
+        fields[p.name] = (
+            pydantic_type,
+            Field(default=default, description=description),
+        )
+    return create_model(
+        f"{function.catalog_name}__{function.schema_name}__{function.name}__params",
+        **fields,  # type: ignore
+    )
+
+
+def _get_tool_name(function: "FunctionInfo") -> str:
+    tool_name = f"{function.catalog_name}__{function.schema_name}__{function.name}"[
+        -64:
+    ]
+    return tool_name
+
+
+def _get_default_workspace_client() -> "WorkspaceClient":
+    try:
+        from databricks.sdk import WorkspaceClient
+    except ImportError as e:
+        raise ImportError(
+            "Could not import databricks-sdk python package. "
+            "Please install it with `pip install databricks-sdk`."
+        ) from e
+    return WorkspaceClient()
+
+
+class UCFunctionToolkit(BaseToolkit):
+    warehouse_id: str = Field(
+        description="The ID of a Databricks SQL Warehouse to execute functions."
+    )
+
+    workspace_client: "WorkspaceClient" = Field(
+        default_factory=_get_default_workspace_client,
+        description="Databricks workspace client.",
+    )
+
+    tools: Dict[str, BaseTool] = Field(default_factory=dict)
+
+    class Config:
+        arbitrary_types_allowed = True
+
+    def include(self, *function_names: str, **kwargs: Any) -> Self:
+        """
+        Includes UC functions to the toolkit.
+
+        Args:
+            functions: A list of UC function names in the format
+                "catalog_name.schema_name.function_name" or
+                "catalog_name.schema_name.*".
+                If the function name ends with ".*",
+                all functions in the schema will be added.
+            kwargs: Extra arguments to pass to StructuredTool, e.g., `return_direct`.
+        """
+        for name in function_names:
+            if name.endswith(".*"):
+                catalog_name, schema_name = name[:-2].split(".")
+                # TODO: handle pagination, warn and truncate if too many
+                functions = self.workspace_client.functions.list(
+                    catalog_name=catalog_name, schema_name=schema_name
+                )
+                for f in functions:
+                    assert f.full_name is not None
+                    self.include(f.full_name, **kwargs)
+            else:
+                if name not in self.tools:
+                    self.tools[name] = self._make_tool(name, **kwargs)
+        return self
+
+    def _make_tool(self, function_name: str, **kwargs: Any) -> BaseTool:
+        function = self.workspace_client.functions.get(function_name)
+        name = _get_tool_name(function)
+        description = function.comment or ""
+        args_schema = _generate_args_schema(function)
+
+        def func(*args: Any, **kwargs: Any) -> str:
+            # TODO: We expect all named args and ignore args.
+            # Non-empty args show up when the function has no parameters.
+            args_json = json.loads(json.dumps(kwargs, default=str))
+            result = execute_function(
+                ws=self.workspace_client,
+                warehouse_id=self.warehouse_id,
+                function=function,
+                parameters=args_json,
+            )
+            return result.to_json()
+
+        return StructuredTool(
+            name=name,
+            description=description,
+            args_schema=args_schema,
+            func=func,
+            **kwargs,
+        )
+
+    def get_tools(self) -> List[BaseTool]:
+        return list(self.tools.values())