LangChain v0.1.0 (jupyterlab#572)

* implement ProviderMetaclass

* add ProviderMetaclass test

* bump langchain to v0.1.0

* migrate import statements deprecated by langchain v0.1

* lint

packages/jupyter-ai-magics/jupyter_ai_magics/embedding_providers.py

@@ -1,4 +1,3 @@
-import os
 from typing import ClassVar, List
 
 from jupyter_ai_magics.providers import (
@@ -8,15 +7,15 @@
     Field,
     MultiEnvAuthStrategy,
 )
-from langchain.embeddings import (
+from langchain.pydantic_v1 import BaseModel, Extra
+from langchain_community.embeddings import (
     BedrockEmbeddings,
     CohereEmbeddings,
     GPT4AllEmbeddings,
     HuggingFaceHubEmbeddings,
     OpenAIEmbeddings,
     QianfanEmbeddingsEndpoint,
 )
-from langchain.pydantic_v1 import BaseModel, Extra
 
 
 class BaseEmbeddingsProvider(BaseModel):

packages/jupyter-ai-magics/jupyter_ai_magics/providers.py

@@ -5,45 +5,41 @@
 import io
 import json
 from concurrent.futures import ThreadPoolExecutor
-from typing import (
-    Any,
-    ClassVar,
-    Coroutine,
-    Dict,
-    List,
-    Literal,
-    Mapping,
-    Optional,
-    Union,
-)
+from typing import Any, ClassVar, Coroutine, Dict, List, Literal, Optional, Union
 
 from jsonpath_ng import parse
-from langchain.chat_models import (
+from langchain.chat_models.base import BaseChatModel
+from langchain.llms.sagemaker_endpoint import LLMContentHandler
+from langchain.llms.utils import enforce_stop_tokens
+from langchain.prompts import PromptTemplate
+from langchain.pydantic_v1 import BaseModel, Extra, root_validator
+from langchain.schema import LLMResult
+from langchain.utils import get_from_dict_or_env
+from langchain_community.chat_models import (
     AzureChatOpenAI,
     BedrockChat,
     ChatAnthropic,
+    ChatOpenAI,
     QianfanChatEndpoint,
 )
-from langchain.chat_models.base import BaseChatModel
-from langchain.llms import (
+from langchain_community.llms import (
     AI21,
     Anthropic,
     Bedrock,
     Cohere,
     GPT4All,
     HuggingFaceHub,
     OpenAI,
-    OpenAIChat,
-    QianfanLLMEndpoint,
     SagemakerEndpoint,
 )
-from langchain.llms.sagemaker_endpoint import LLMContentHandler
-from langchain.llms.utils import enforce_stop_tokens
-from langchain.prompts import PromptTemplate
-from langchain.pydantic_v1 import BaseModel, Extra, root_validator
-from langchain.schema import LLMResult
-from langchain.utils import get_from_dict_or_env
-from langchain_community.chat_models import ChatOpenAI
+
+# this is necessary because `langchain.pydantic_v1.main` does not include
+# `ModelMetaclass`, as it is not listed in `__all__` by the `pydantic.main`
+# subpackage.
+try:
+    from pydantic.v1.main import ModelMetaclass
+except:
+    from pydantic.main import ModelMetaclass
 
 
 class EnvAuthStrategy(BaseModel):
@@ -99,7 +95,34 @@ class IntegerField(BaseModel):
 Field = Union[TextField, MultilineTextField, IntegerField]
 
 
-class BaseProvider(BaseModel):
+class ProviderMetaclass(ModelMetaclass):
+    """
+    A metaclass that ensures all class attributes defined inline within the
+    class definition are accessible and included in `Class.__dict__`.
+
+    This is necessary because Pydantic drops any ClassVars that are defined as
+    an instance field by a parent class, even if they are defined inline within
+    the class definition. We encountered this case when `langchain` added a
+    `name` attribute to a parent class shared by all `Provider`s, which caused
+    `Provider.name` to be inaccessible. See #558 for more info.
+    """
+
+    def __new__(mcs, name, bases, namespace, **kwargs):
+        cls = super().__new__(mcs, name, bases, namespace, **kwargs)
+        for key in namespace:
+            # skip private class attributes
+            if key.startswith("_"):
+                continue
+            # skip class attributes already listed in `cls.__dict__`
+            if key in cls.__dict__:
+                continue
+
+            setattr(cls, key, namespace[key])
+
+        return cls
+
+
+class BaseProvider(BaseModel, metaclass=ProviderMetaclass):
     #
     # pydantic config
     #

packages/jupyter-ai-magics/jupyter_ai_magics/tests/test_provider_metaclass.py

@@ -0,0 +1,26 @@
+from typing import ClassVar, Optional
+
+from langchain.pydantic_v1 import BaseModel
+
+from ..providers import ProviderMetaclass
+
+
+def test_provider_metaclass():
+    """
+    Asserts that the metaclass prevents class attributes from being omitted due
+    to parent classes defining an instance field of the same name.
+
+    You can reproduce the original issue by removing the
+    `metaclass=ProviderMetaclass` argument from the definition of `Child`.
+    """
+
+    class Parent(BaseModel):
+        test: Optional[str]
+
+    class Base(BaseModel):
+        test: ClassVar[str]
+
+    class Child(Base, Parent, metaclass=ProviderMetaclass):
+        test: ClassVar[str] = "expected"
+
+    assert Child.test == "expected"

packages/jupyter-ai-magics/pyproject.toml

@@ -23,24 +23,16 @@ dynamic = ["version", "description", "authors", "urls", "keywords"]
 dependencies = [
     "ipython",
     "importlib_metadata>=5.2.0",
-    "langchain==0.0.350",
-    "langchain-core>=0.1.0,<0.1.4",
+    "langchain>=0.1.0,<0.2.0",
     "typing_extensions>=4.5.0",
     "click~=8.0",
     "jsonpath-ng>=1.5.3,<2",
 ]
 
 [project.optional-dependencies]
-dev = [
-    "pre-commit>=3.3.3,<4"
-]
+dev = ["pre-commit>=3.3.3,<4"]
 
-test = [
-    "coverage",
-    "pytest",
-    "pytest-asyncio",
-    "pytest-cov"
-]
+test = ["coverage", "pytest", "pytest-asyncio", "pytest-cov"]
 
 all = [
     "ai21",
@@ -52,7 +44,7 @@ all = [
     "pillow",
     "openai~=1.6.1",
     "boto3",
-    "qianfan"
+    "qianfan",
 ]
 
 [project.entry-points."jupyter_ai.model_providers"]

packages/jupyter-ai/jupyter_ai/chat_handlers/learn.py

@@ -1,10 +1,9 @@
 import argparse
 import json
 import os
-from typing import Any, Awaitable, Coroutine, List, Optional, Tuple
+from typing import Any, Coroutine, List, Optional, Tuple
 
 from dask.distributed import Client as DaskClient
-from jupyter_ai.config_manager import ConfigManager
 from jupyter_ai.document_loaders.directory import get_embeddings, split
 from jupyter_ai.document_loaders.splitter import ExtensionSplitter, NotebookSplitter
 from jupyter_ai.models import (
@@ -22,7 +21,7 @@
     PythonCodeTextSplitter,
     RecursiveCharacterTextSplitter,
 )
-from langchain.vectorstores import FAISS
+from langchain_community.vectorstores import FAISS
 
 from .base import BaseChatHandler, SlashCommandRoutingType
 
@@ -141,7 +140,7 @@ def _build_list_response(self):
     async def learn_dir(
         self, path: str, chunk_size: int, chunk_overlap: int, all_files: bool
     ):
-        dask_client = await self.dask_client_future
+        dask_client: DaskClient = await self.dask_client_future
         splitter_kwargs = {"chunk_size": chunk_size, "chunk_overlap": chunk_overlap}
         splitters = {
             ".py": PythonCodeTextSplitter(**splitter_kwargs),

packages/jupyter-ai/pyproject.toml

@@ -26,8 +26,6 @@ dependencies = [
     "jupyterlab~=4.0",
     "aiosqlite>=0.18",
     "importlib_metadata>=5.2.0",
-    "langchain==0.0.350",
-    "langchain-core>=0.1.0,<0.1.4",
     "tiktoken",                  # required for OpenAIEmbeddings
     "jupyter_ai_magics",
     "dask[distributed]",

pyproject.toml

@@ -34,10 +34,12 @@ include_external_packages = true
 
 [[tool.importlinter.contracts]]
 key = "pydantic"
-name = "Forbidden import of `pydantic` package. Please import from `langchain.pydantic_v1` instead for compatibility with both Pydantic v1 and v2."
+name = "Forbid `pydantic`. (note: Developers should import Pydantic from `langchain.pydantic_v1` instead for compatibility.)"
 type = "forbidden"
 source_modules = ["jupyter_ai", "jupyter_ai_magics"]
 forbidden_modules = ["pydantic"]
+# TODO: get `langchain` to export `ModelMetaclass` to avoid needing this statement
+ignore_imports = ["jupyter_ai_magics.providers -> pydantic"]
 
 [tool.pytest.ini_options]
 addopts = "--ignore packages/jupyter-ai-module-cookiecutter"