Merge branch 'python_custom_node_component' of github.com:logspace-ai/langflow into python_custom_node_component

This commit is contained in:
Lucas Oliveira 2023-07-28 14:49:11 -03:00
commit 1ea05b3584
92 changed files with 4582 additions and 1520 deletions

View file

@ -1,7 +1,7 @@
from importlib import metadata
from langflow.cache import cache_manager
from langflow.processing.process import load_flow_from_json
from langflow.utils.types import Prompt
from langflow.interface.custom.custom_component import CustomComponent
try:
__version__ = metadata.version(__package__)
@ -10,4 +10,4 @@ except metadata.PackageNotFoundError:
__version__ = ""
del metadata # optional, avoids polluting the results of dir(__package__)
__all__ = ["load_flow_from_json", "cache_manager", "Prompt"]
__all__ = ["load_flow_from_json", "cache_manager", "CustomComponent"]

View file

@ -2,8 +2,9 @@ import ast
import inspect
import traceback
from typing import Dict, Any, Type, Union
from typing import Dict, Any, List, Type, Union
from fastapi import HTTPException
from langflow.interface.custom.schema import CallableCodeDetails, ClassCodeDetails
class CodeSyntaxError(HTTPException):
@ -54,13 +55,13 @@ class CodeParser:
return tree
def parse_node(self, node: ast.AST) -> None:
def parse_node(self, node: Union[ast.stmt, ast.AST]) -> None:
"""
Parses an AST node and updates the data
dictionary with the relevant information.
"""
if handler := self.handlers.get(type(node)):
handler(node)
if handler := self.handlers.get(type(node)): # type: ignore
handler(node) # type: ignore
def parse_imports(self, node: Union[ast.Import, ast.ImportFrom]) -> None:
"""
@ -92,27 +93,73 @@ class CodeParser:
"""
Extracts details from a single function or method node.
"""
func = {
"name": node.name,
"doc": ast.get_docstring(node),
"args": [],
"body": [],
"return_type": ast.unparse(node.returns) if node.returns else None,
}
func = CallableCodeDetails(
name=node.name,
doc=ast.get_docstring(node),
args=[],
body=[],
return_type=ast.unparse(node.returns) if node.returns else None,
)
# Handle positional arguments with default values
defaults = [None] * (len(node.args.args) - len(node.args.defaults)) + [
ast.unparse(default) if default else None for default in node.args.defaults
func.args = self.parse_function_args(node)
func.body = self.parse_function_body(node)
return func.dict()
def parse_function_args(self, node: ast.FunctionDef) -> List[Dict[str, Any]]:
"""
Parses the arguments of a function or method node.
"""
args = []
args += self.parse_positional_args(node)
args += self.parse_varargs(node)
args += self.parse_keyword_args(node)
args += self.parse_kwargs(node)
return args
def parse_positional_args(self, node: ast.FunctionDef) -> List[Dict[str, Any]]:
"""
Parses the positional arguments of a function or method node.
"""
num_args = len(node.args.args)
num_defaults = len(node.args.defaults)
num_missing_defaults = num_args - num_defaults
missing_defaults = [None] * num_missing_defaults
default_values = [
ast.unparse(default).strip("'") if default else None
for default in node.args.defaults
]
# Now check all default values to see if there
# are any "None" values in the middle
default_values = [
None if value == "None" else value for value in default_values
]
for arg, default in zip(node.args.args, defaults):
func["args"].append(self.parse_arg(arg, default))
defaults = missing_defaults + default_values
args = [
self.parse_arg(arg, default)
for arg, default in zip(node.args.args, defaults)
]
return args
def parse_varargs(self, node: ast.FunctionDef) -> List[Dict[str, Any]]:
"""
Parses the *args argument of a function or method node.
"""
args = []
# Handle *args
if node.args.vararg:
func["args"].append(self.parse_arg(node.args.vararg, None))
args.append(self.parse_arg(node.args.vararg, None))
# Handle keyword-only arguments with default values
return args
def parse_keyword_args(self, node: ast.FunctionDef) -> List[Dict[str, Any]]:
"""
Parses the keyword-only arguments of a function or method node.
"""
kw_defaults = [None] * (
len(node.args.kwonlyargs) - len(node.args.kw_defaults)
) + [
@ -120,16 +167,28 @@ class CodeParser:
for default in node.args.kw_defaults
]
for arg, default in zip(node.args.kwonlyargs, kw_defaults):
func["args"].append(self.parse_arg(arg, default))
args = [
self.parse_arg(arg, default)
for arg, default in zip(node.args.kwonlyargs, kw_defaults)
]
return args
def parse_kwargs(self, node: ast.FunctionDef) -> List[Dict[str, Any]]:
"""
Parses the **kwargs argument of a function or method node.
"""
args = []
# Handle **kwargs
if node.args.kwarg:
func["args"].append(self.parse_arg(node.args.kwarg, None))
args.append(self.parse_arg(node.args.kwarg, None))
for line in node.body:
func["body"].append(ast.unparse(line))
return func
return args
def parse_function_body(self, node: ast.FunctionDef) -> List[str]:
"""
Parses the body of a function or method node.
"""
return [ast.unparse(line) for line in node.body]
def parse_assign(self, stmt):
"""
@ -164,29 +223,31 @@ class CodeParser:
"""
Extracts "classes" from the code, including inheritance and init methods.
"""
class_dict = {
"name": node.name,
"doc": ast.get_docstring(node),
"bases": [ast.unparse(base) for base in node.bases],
"attributes": [],
"methods": [],
}
class_details = ClassCodeDetails(
name=node.name,
doc=ast.get_docstring(node),
bases=[ast.unparse(base) for base in node.bases],
attributes=[],
methods=[],
init=None,
)
for stmt in node.body:
if isinstance(stmt, ast.Assign):
if attr := self.parse_assign(stmt):
class_dict["attributes"].append(attr)
class_details.attributes.append(attr)
elif isinstance(stmt, ast.AnnAssign):
if attr := self.parse_ann_assign(stmt):
class_dict["attributes"].append(attr)
class_details.attributes.append(attr)
elif isinstance(stmt, ast.FunctionDef):
method, is_init = self.parse_function_def(stmt)
if is_init:
class_dict["init"] = method
class_details.init = method
else:
class_dict["methods"].append(method)
class_details.methods.append(method)
self.data["classes"].append(class_dict)
self.data["classes"].append(class_details.dict())
def parse_global_vars(self, node: ast.Assign) -> None:
"""

View file

@ -1,4 +1,5 @@
import ast
from typing import Optional
from pydantic import BaseModel
from fastapi import HTTPException
@ -20,7 +21,7 @@ class Component(BaseModel):
"The name of the entrypoint function must be provided."
)
code: str
code: Optional[str]
function_entrypoint_name = "build"
field_config: dict = {}

View file

@ -20,13 +20,22 @@ LANGCHAIN_BASE_TYPES = {
"VectorStore": VectorStore,
"Embeddings": Embeddings,
"BaseRetriever": BaseRetriever,
}
# Langchain base types plus Python base types
CUSTOM_COMPONENT_SUPPORTED_TYPES = {
**LANGCHAIN_BASE_TYPES,
"str": str,
"int": int,
"float": float,
"bool": bool,
"list": list,
"dict": dict,
}
DEFAULT_CUSTOM_COMPONENT_CODE = """
from langflow import Prompt
from langflow.interface.custom.custom_component import CustomComponent
from langflow import CustomComponent
from langchain.llms.base import BaseLLM
from langchain.chains import LLMChain
@ -38,11 +47,12 @@ import requests
class YourComponent(CustomComponent):
display_name: str = "Your Component"
description: str = "Your description"
field_config = { "url": { "multiline": True, "required": True } }
def build(self, url: str, llm: BaseLLM, template: Prompt) -> Document:
def build_config(self):
return { "url": { "multiline": True, "required": True } }
def build(self, url: str, llm: BaseLLM, prompt: PromptTemplate) -> Document:
response = requests.get(url)
prompt = PromptTemplate.from_template(template)
chain = LLMChain(llm=llm, prompt=prompt)
result = chain.run(response.text[:300])
return Document(page_content=str(result))

View file

@ -1,11 +1,10 @@
from typing import Callable, Optional
from typing import Any, Callable, List, Optional
from fastapi import HTTPException
from langflow.interface.custom.constants import LANGCHAIN_BASE_TYPES
from langflow.interface.custom.constants import CUSTOM_COMPONENT_SUPPORTED_TYPES
from langflow.interface.custom.component import Component
from langflow.utils import validate
from uuid import UUID
from langflow.database.base import session_getter
from langflow.database.models.flow import Flow
from pydantic import Extra
@ -17,14 +16,14 @@ class CustomComponent(Component, extra=Extra.allow):
code_class_base_inheritance = "CustomComponent"
function_entrypoint_name = "build"
function: Optional[Callable] = None
return_type_valid_list = list(LANGCHAIN_BASE_TYPES.keys())
return_type_valid_list = list(CUSTOM_COMPONENT_SUPPORTED_TYPES.keys())
repr_value: Optional[str] = ""
def __init__(self, **data):
super().__init__(**data)
def custom_repr(self):
return self.repr_value
return str(self.repr_value)
def build_config(self):
return self.field_config
@ -44,13 +43,15 @@ class CustomComponent(Component, extra=Extra.allow):
return True
def is_check_valid(self) -> bool:
return self._class_template_validation(self.code)
return self._class_template_validation(self.code) if self.code else False
def get_code_tree(self, code: str):
return super().get_code_tree(code)
@property
def get_function_entrypoint_args(self) -> str:
if not self.code:
return ""
tree = self.get_code_tree(self.code)
component_classes = [
@ -78,6 +79,8 @@ class CustomComponent(Component, extra=Extra.allow):
@property
def get_function_entrypoint_return_type(self) -> str:
if not self.code:
return ""
tree = self.get_code_tree(self.code)
component_classes = [
@ -138,16 +141,19 @@ class CustomComponent(Component, extra=Extra.allow):
def get_function(self):
return validate.create_function(self.code, self.function_entrypoint_name)
def load_flow(self, flow_id: UUID = None):
def load_flow(self, flow_id: str, tweaks: Optional[dict] = None) -> Any:
from langflow.processing.process import build_sorted_vertices_with_caching
from langflow.processing.process import process_tweaks
with session_getter() as session:
data_graph = flow.data if (flow := session.get(Flow, flow_id)) else None
if not data_graph:
graph_data = flow.data if (flow := session.get(Flow, flow_id)) else None
if not graph_data:
raise ValueError(f"Flow {flow_id} not found")
return build_sorted_vertices_with_caching(data_graph)
if tweaks:
graph_data = process_tweaks(graph_data=graph_data, tweaks=tweaks)
return build_sorted_vertices_with_caching(graph_data)
def list_flows(self):
def list_flows(self) -> List[Flow]:
with session_getter() as session:
flows = session.query(Flow).all()
return flows

View file

@ -0,0 +1,29 @@
from pydantic import BaseModel, Field
from typing import Optional
class ClassCodeDetails(BaseModel):
"""
A dataclass for storing details about a class.
"""
name: str
doc: Optional[str]
bases: list
attributes: list
methods: list
init: Optional[dict] = Field(default_factory=dict)
class CallableCodeDetails(BaseModel):
"""
A dataclass for storing details about a callable.
"""
name: str
doc: Optional[str]
args: list
body: list
return_type: Optional[str]

View file

@ -1,17 +1,15 @@
import contextlib
import json
from typing import Any, Callable, Dict, List, Sequence, Type
from typing import Any, Callable, Dict, Sequence, Type
from langchain.agents import ZeroShotAgent
from langchain.agents import agent as agent_module
from langchain.agents.agent import AgentExecutor
from langchain.agents.agent_toolkits.base import BaseToolkit
from langchain.agents.tools import BaseTool
from langflow.interface.initialize.llm import initialize_vertexai
from langflow.interface.initialize.utils import handle_format_kwargs, handle_node_type
from langflow.interface.initialize.vector_store import vecstore_initializer
from langchain.schema import Document, BaseOutputParser
from pydantic import ValidationError
from langflow.interface.importing.utils import (
@ -212,68 +210,8 @@ def instantiate_agent(node_type, class_object: Type[agent_module.Agent], params:
def instantiate_prompt(node_type, class_object, params: Dict):
if node_type == "ZeroShotPrompt":
if "tools" not in params:
params["tools"] = []
return ZeroShotAgent.create_prompt(**params)
elif "MessagePromptTemplate" in node_type:
# Then we only need the template
from_template_params = {
"template": params.pop("prompt", params.pop("template", ""))
}
if not from_template_params.get("template"):
raise ValueError("Prompt template is required")
prompt = class_object.from_template(**from_template_params)
elif node_type == "ChatPromptTemplate":
prompt = class_object.from_messages(**params)
else:
prompt = class_object(**params)
format_kwargs: Dict[str, Any] = {}
for input_variable in prompt.input_variables:
if input_variable in params:
variable = params[input_variable]
if isinstance(variable, str):
format_kwargs[input_variable] = variable
elif isinstance(variable, BaseOutputParser) and hasattr(
variable, "get_format_instructions"
):
format_kwargs[input_variable] = variable.get_format_instructions()
elif isinstance(variable, List) and all(
isinstance(item, Document) for item in variable
):
# Format document to contain page_content and metadata
# as one string separated by a newline
if len(variable) > 1:
content = "\n".join(
[item.page_content for item in variable if item.page_content]
)
else:
content = variable[0].page_content
# content could be a json list of strings
with contextlib.suppress(json.JSONDecodeError):
content = json.loads(content)
if isinstance(content, list):
content = ",".join([str(item) for item in content])
format_kwargs[input_variable] = content
# handle_keys will be a list but it does not exist yet
# so we need to create it
if (
isinstance(variable, List)
and all(isinstance(item, Document) for item in variable)
) or (
isinstance(variable, BaseOutputParser)
and hasattr(variable, "get_format_instructions")
):
if "handle_keys" not in format_kwargs:
format_kwargs["handle_keys"] = []
# Add the handle_keys to the list
format_kwargs["handle_keys"].append(input_variable)
params, prompt = handle_node_type(node_type, class_object, params)
format_kwargs = handle_format_kwargs(prompt, params)
return prompt, format_kwargs

View file

@ -0,0 +1,103 @@
import contextlib
import json
from typing import Any, Dict, List
from langchain.agents import ZeroShotAgent
from langchain.schema import Document, BaseOutputParser
def handle_node_type(node_type, class_object, params: Dict):
if node_type == "ZeroShotPrompt":
params = check_tools_in_params(params)
prompt = ZeroShotAgent.create_prompt(**params)
elif "MessagePromptTemplate" in node_type:
prompt = instantiate_from_template(class_object, params)
elif node_type == "ChatPromptTemplate":
prompt = class_object.from_messages(**params)
else:
prompt = class_object(**params)
return params, prompt
def check_tools_in_params(params: Dict):
if "tools" not in params:
params["tools"] = []
return params
def instantiate_from_template(class_object, params: Dict):
from_template_params = {
"template": params.pop("prompt", params.pop("template", ""))
}
if not from_template_params.get("template"):
raise ValueError("Prompt template is required")
return class_object.from_template(**from_template_params)
def handle_format_kwargs(prompt, params: Dict):
format_kwargs: Dict[str, Any] = {}
for input_variable in prompt.input_variables:
if input_variable in params:
format_kwargs = handle_variable(params, input_variable, format_kwargs)
return format_kwargs
def handle_variable(params: Dict, input_variable: str, format_kwargs: Dict):
variable = params[input_variable]
if isinstance(variable, str):
format_kwargs[input_variable] = variable
elif isinstance(variable, BaseOutputParser) and hasattr(
variable, "get_format_instructions"
):
format_kwargs[input_variable] = variable.get_format_instructions()
elif is_instance_of_list_or_document(variable):
format_kwargs = format_document(variable, input_variable, format_kwargs)
if needs_handle_keys(variable):
format_kwargs = add_handle_keys(input_variable, format_kwargs)
return format_kwargs
def is_instance_of_list_or_document(variable):
return (
isinstance(variable, List)
and all(isinstance(item, Document) for item in variable)
or isinstance(variable, Document)
)
def format_document(variable, input_variable: str, format_kwargs: Dict):
variable = variable if isinstance(variable, List) else [variable]
content = format_content(variable)
format_kwargs[input_variable] = content
return format_kwargs
def format_content(variable):
if len(variable) > 1:
return "\n".join([item.page_content for item in variable if item.page_content])
content = variable[0].page_content
return try_to_load_json(content)
def try_to_load_json(content):
with contextlib.suppress(json.JSONDecodeError):
content = json.loads(content)
if isinstance(content, list):
content = ",".join([str(item) for item in content])
return content
def needs_handle_keys(variable):
return is_instance_of_list_or_document(variable) or (
isinstance(variable, BaseOutputParser)
and hasattr(variable, "get_format_instructions")
)
def add_handle_keys(input_variable: str, format_kwargs: Dict):
if "handle_keys" not in format_kwargs:
format_kwargs["handle_keys"] = []
format_kwargs["handle_keys"].append(input_variable)
return format_kwargs

View file

@ -55,7 +55,7 @@ TOOL_INPUTS = {
show=True,
value="",
suffixes=[".json", ".yaml", ".yml"],
fileTypes=["json", "yaml", "yml"],
file_types=["json", "yaml", "yml"],
),
}

View file

@ -48,29 +48,3 @@ class PythonFunctionTool(Function, Tool):
class PythonFunction(Function):
code: str
class CustomComponent_old(BaseModel):
code: str
function: Optional[Callable] = None
imports: Optional[str] = None
# Eval code and store the class
def __init__(self, **data):
super().__init__(**data)
# Validate the Class code
@validator("code")
def validate_func(cls, v):
try:
validate.eval_function(v)
except Exception as e:
raise e
return v
def get_function(self):
"""Get the function"""
function_name = validate.extract_function_name(self.code)
return validate.create_function(self.code, function_name)

View file

@ -1,6 +1,6 @@
from langflow.interface.agents.base import agent_creator
from langflow.interface.chains.base import chain_creator
from langflow.interface.custom.constants import LANGCHAIN_BASE_TYPES
from langflow.interface.custom.constants import CUSTOM_COMPONENT_SUPPORTED_TYPES
from langflow.interface.document_loaders.base import documentloader_creator
from langflow.interface.embeddings.base import embedding_creator
from langflow.interface.importing.utils import get_function_custom
@ -98,6 +98,13 @@ def add_new_custom_field(
display_name = field_config.pop("display_name", field_name)
field_type = field_config.pop("field_type", field_type)
field_type = process_type(field_type)
field_value = field_config.pop("value", field_value)
field_advanced = field_config.pop("advanced", False)
# If options is a list, then it's a dropdown
# If options is None, then it's a list of strings
is_list = isinstance(field_config.get("options"), list)
field_config["is_list"] = is_list or field_config.get("is_list", False)
if "name" in field_config:
warnings.warn(
@ -114,7 +121,7 @@ def add_new_custom_field(
value=field_value,
show=True,
required=required,
advanced=False,
advanced=field_advanced,
placeholder=placeholder,
display_name=display_name,
**field_config,
@ -126,8 +133,9 @@ def add_new_custom_field(
# TODO: Move to correct place
def add_code_field(template, raw_code):
def add_code_field(template, raw_code, field_config):
# Field with the Python code to allow update
code_field = {
"code": {
"dynamic": True,
@ -138,7 +146,7 @@ def add_code_field(template, raw_code):
"value": raw_code,
"password": False,
"name": "code",
"advanced": False,
"advanced": field_config.pop("advanced", False),
"type": "code",
"list": False,
}
@ -183,22 +191,30 @@ def update_display_name_and_description(frontend_node, template_config):
frontend_node["description"] = template_config["description"]
def build_field_config(custom_component):
def build_field_config(custom_component: CustomComponent):
"""Build the field configuration for a custom component"""
try:
custom_class = get_function_custom(custom_component.code)
return custom_class().build_config()
except Exception as exc:
logger.error(f"Error while building field config: {exc}")
logger.error(f"Error while getting custom function: {str(exc)}")
return {}
try:
return custom_class().build_config()
except Exception as exc:
logger.error(f"Error while building field config: {str(exc)}")
return {}
def add_extra_fields(frontend_node, field_config, function_args):
"""Add extra fields to the frontend node"""
if function_args is None:
if function_args is None or function_args == "":
return
# sort function_args which is a list of dicts
function_args.sort(key=lambda x: x["name"])
for extra_field in function_args:
if "name" not in extra_field or extra_field["name"] == "self":
continue
@ -232,19 +248,19 @@ def get_field_properties(extra_field):
def add_base_classes(frontend_node, return_type):
"""Add base classes to the frontend node"""
if return_type not in LANGCHAIN_BASE_TYPES or return_type is None:
if return_type not in CUSTOM_COMPONENT_SUPPORTED_TYPES or return_type is None:
raise HTTPException(
status_code=400,
detail={
"error": (
"Invalid return type should be one of: "
f"{list(LANGCHAIN_BASE_TYPES.keys())}"
f"{list(CUSTOM_COMPONENT_SUPPORTED_TYPES.keys())}"
),
"traceback": traceback.format_exc(),
},
)
return_type_instance = LANGCHAIN_BASE_TYPES.get(return_type)
return_type_instance = CUSTOM_COMPONENT_SUPPORTED_TYPES.get(return_type)
base_classes = get_base_classes(return_type_instance)
for base_class in base_classes:
@ -268,7 +284,9 @@ def build_langchain_template_custom_component(custom_component: CustomComponent)
frontend_node, field_config, custom_component.get_function_entrypoint_args
)
frontend_node = add_code_field(frontend_node, custom_component.code)
frontend_node = add_code_field(
frontend_node, custom_component.code, field_config.get("code", {})
)
add_base_classes(
frontend_node, custom_component.get_function_entrypoint_return_type
@ -287,8 +305,8 @@ def load_files_from_path(path: str):
def build_and_validate_all_files(reader, file_list):
"""Build and validate all files"""
data = reader.build_component_menu_list(file_list)
valid_components = reader.filter_loaded_components(data=data, with_errors=False)
valid_components = reader.filter_loaded_components(data=data, with_errors=False)
invalid_components = reader.filter_loaded_components(data=data, with_errors=True)
return valid_components, invalid_components
@ -341,12 +359,15 @@ def build_invalid_menu(invalid_components):
.get(type(CustomComponent()).__name__)
)
component_template["error"] = component.get("error", None)
component_template.get("template").get("code")["value"] = component_code
invalid_menu[menu_name][component_name] = component_template
except Exception as exc:
logger.error(f"Error while creating custom component: {exc}")
logger.error(
f"Error while creating custom component [{component_name}]: {str(exc)}"
)
return invalid_menu

View file

@ -145,7 +145,7 @@ class CSVAgentNode(FrontendNode):
name="path",
value="",
suffixes=[".csv"],
fileTypes=["csv"],
file_types=["csv"],
),
TemplateField(
field_type="BaseLanguageModel",

View file

@ -53,6 +53,7 @@ class FrontendNode(BaseModel):
output_types: List[str] = []
field_formatters: FieldFormatters = Field(default_factory=FieldFormatters)
beta: bool = False
error: Optional[str] = None
# field formatters is an instance attribute but it is not used in the class
# so we need to create a method to get it
@ -85,6 +86,7 @@ class FrontendNode(BaseModel):
"output_types": self.output_types,
"documentation": self.documentation,
"beta": self.beta,
"error": self.error,
},
}

View file

@ -8,6 +8,7 @@ FORCE_SHOW_FIELDS = [
"headers",
"max_value_length",
"max_tokens",
"google_cse_id",
]
DEFAULT_PROMPT = """

View file

@ -14,7 +14,7 @@ def build_file_field(
name=name,
value="",
suffixes=suffixes,
fileTypes=fileTypes,
file_types=fileTypes,
)

View file

@ -19,7 +19,7 @@ class LLMFrontendNode(FrontendNode):
name="credentials",
value="",
suffixes=[".json"],
fileTypes=["json"],
file_types=["json"],
)
)

View file

@ -9,7 +9,7 @@ from docstring_parser import parse # type: ignore
from langflow.template.frontend_node.constants import FORCE_SHOW_FIELDS
from langflow.utils import constants
from langflow.utils.logger import logger
from multiprocess import cpu_count
from multiprocess import cpu_count # type: ignore
def build_template_from_function(
@ -301,13 +301,15 @@ def get_type(value: Any) -> Union[str, type]:
return _type if isinstance(_type, str) else _type.__name__
def remove_optional_wrapper(_type: str) -> str:
def remove_optional_wrapper(_type: Union[str, type]) -> str:
"""
Removes the 'Optional' wrapper from the type string.
Returns:
The type string with the 'Optional' wrapper removed.
"""
if isinstance(_type, type):
_type = str(_type)
if "Optional" in _type:
_type = _type.replace("Optional[", "")[:-1]