LangChain Compability Restored (#442)

* langchain compatible llminterfacev2 interface created

* langchain compatible methods added to openai

* broken test cases sorted

* brand new tests for llm interface v2

* vertex ai started to support llm interface v2

* brand new test cases added

* invoke with tools method is not mandatory

* llminterfacev2 support added to ollama

* llm interface v2 supported in mistralai

* llm interface v2 support added to cohere

* llm interface v2 support added to anthropic

* more tests for llminterfacev2

* mypy fixes

* test_openai_llm possibly failed because of import in ci cd

* Attempt CI/CD-compatible async mock for OpenAILLM tests

* Attempt CI/CD-safe async mock for OpenAILLM v1 test

* restoring graphrag e2e tests for v1

* existing e2e test for graphrag sorted

* create message history only if langchain compatible branch

* typo in docstring

* avoid repeated langchain compatible check code

* avoid to create vector idx and fulltext with same property

* avoid to create vector idx and fulltext with same property

* make is instance of langchain check prettier

* use params from invoke if available similar to LC

* define kwargs for invoke in the interface

* fixing outer scope definition warning

* arg names replaced with LC arg names

* some additional docstring for new input args

* docstring updated

* use (a)invokev2 function names instead of brand new

* use (a)invoke_v1 function name instead of legacy invoke

* keep rate limit handlers for v2 functions

* use warning for deprecated llm interface v1

* resolve conflict after Alex's recent uv change

* formatted

* mypy problems sorted

* initialize llms with llm interface v2

* invoke with tools dropped in llm interface v2

* revert the behaviour back for return_context

* e2e tests updated for restored behaviour for return_context

* llm can also be lc object, added any for this

* users initialize their own llm objects, so we don't have to make llm interface's init same as lc

* auto format after manual conflict resolution

Author: serengil

src/neo4j_graphrag/generation/graphrag.py

@@ -12,29 +12,35 @@
 #  WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
 #  See the License for the specific language governing permissions and
 #  limitations under the License.
-from __future__ import annotations
 
+# built-in dependencies
+from __future__ import annotations
 import logging
 import warnings
 from typing import Any, List, Optional, Union
 
+# 3rd party dependencies
 from pydantic import ValidationError
 
+# project dependencies
 from neo4j_graphrag.exceptions import (
     RagInitializationError,
     SearchValidationError,
 )
 from neo4j_graphrag.generation.prompts import RagTemplate
 from neo4j_graphrag.generation.types import RagInitModel, RagResultModel, RagSearchModel
-from neo4j_graphrag.llm import LLMInterface
+from neo4j_graphrag.llm import LLMInterface, LLMInterfaceV2
+from neo4j_graphrag.llm.utils import legacy_inputs_to_messages
 from neo4j_graphrag.message_history import MessageHistory
 from neo4j_graphrag.retrievers.base import Retriever
 from neo4j_graphrag.types import LLMMessage, RetrieverResult
 from neo4j_graphrag.utils.logging import prettify
 
+# Set up logger
 logger = logging.getLogger(__name__)
 
 
+# pylint: disable=raise-missing-from
 class GraphRAG:
     """Performs a GraphRAG search using a specific retriever
     and LLM.
@@ -57,8 +63,10 @@ class GraphRAG:
 
     Args:
         retriever (Retriever): The retriever used to find relevant context to pass to the LLM.
-        llm (LLMInterface): The LLM used to generate the answer.
-        prompt_template (RagTemplate): The prompt template that will be formatted with context and user question and passed to the LLM.
+        llm (LLMInterface, LLMInterfaceV2 or LangChain Chat Model): The LLM used to generate
+            the answer.
+        prompt_template (RagTemplate): The prompt template that will be formatted with context and
+            user question and passed to the LLM.
 
     Raises:
         RagInitializationError: If validation of the input arguments fail.
@@ -67,7 +75,7 @@ class GraphRAG:
     def __init__(
         self,
         retriever: Retriever,
-        llm: LLMInterface,
+        llm: Union[LLMInterface, LLMInterfaceV2, Any],
         prompt_template: RagTemplate = RagTemplate(),
     ):
         try:
@@ -93,7 +101,8 @@ def search(
     ) -> RagResultModel:
         """
         .. warning::
-            The default value of 'return_context' will change from 'False' to 'True' in a future version.
+            The default value of 'return_context' will change from 'False'
+                to 'True' in a future version.
 
 
         This method performs a full RAG search:
@@ -104,24 +113,28 @@ def search(
 
         Args:
             query_text (str): The user question.
-            message_history (Optional[Union[List[LLMMessage], MessageHistory]]): A collection previous messages,
-                with each message having a specific role assigned.
+            message_history (Optional[Union[List[LLMMessage], MessageHistory]]): A collection
+                of previous messages, with each message having a specific role assigned.
             examples (str): Examples added to the LLM prompt.
             retriever_config (Optional[dict]): Parameters passed to the retriever.
                 search method; e.g.: top_k
-            return_context (bool): Whether to append the retriever result to the final result (default: False).
-            response_fallback (Optional[str]): If not null, will return this message instead of calling the LLM if context comes back empty.
+            return_context (bool): Whether to append the retriever result to the final result
+                (default: False).
+            response_fallback (Optional[str]): If not null, will return this message instead
+                of calling the LLM if context comes back empty.
 
         Returns:
             RagResultModel: The LLM-generated answer.
 
         """
         if return_context is None:
             warnings.warn(
-                "The default value of 'return_context' will change from 'False' to 'True' in a future version.",
+                "The default value of 'return_context' will change from 'False'"
+                " to 'True' in a future version.",
                 DeprecationWarning,
             )
             return_context = False
+
         try:
             validated_data = RagSearchModel(
                 query_text=query_text,
@@ -145,13 +158,31 @@ def search(
             prompt = self.prompt_template.format(
                 query_text=query_text, context=context, examples=validated_data.examples
             )
-            logger.debug(f"RAG: retriever_result={prettify(retriever_result)}")
-            logger.debug(f"RAG: prompt={prompt}")
-            llm_response = self.llm.invoke(
-                prompt,
-                message_history,
-                system_instruction=self.prompt_template.system_instructions,
-            )
+
+            logger.debug("RAG: retriever_result=%s", prettify(retriever_result))
+            logger.debug("RAG: prompt=%s", prompt)
+
+            if self.is_langchain_compatible():
+                # llm interface v2 or langchain chat model
+                messages = legacy_inputs_to_messages(
+                    prompt=prompt,
+                    message_history=message_history,
+                    system_instruction=self.prompt_template.system_instructions,
+                )
+
+                # langchain chat model compatible invoke
+                llm_response = self.llm.invoke(
+                    input=messages,
+                )
+            elif isinstance(self.llm, LLMInterface):
+                # may have custom LLMs inherited from V1, keep it for backward compatibility
+                llm_response = self.llm.invoke(
+                    input=prompt,
+                    message_history=message_history,
+                    system_instruction=self.prompt_template.system_instructions,
+                )
+            else:
+                raise ValueError(f"Type {type(self.llm)} of LLM is not supported.")
             answer = llm_response.content
         result: dict[str, Any] = {"answer": answer}
         if return_context:
@@ -163,18 +194,47 @@ def _build_query(
         query_text: str,
         message_history: Optional[List[LLMMessage]] = None,
     ) -> str:
-        summary_system_message = "You are a summarization assistant. Summarize the given text in no more than 300 words."
+        """Builds the final query text, incorporating message history if provided."""
+        summary_system_message = (
+            "You are a summarization assistant. "
+            "Summarize the given text in no more than 300 words."
+        )
         if message_history:
             summarization_prompt = self._chat_summary_prompt(
                 message_history=message_history
             )
-            summary = self.llm.invoke(
-                input=summarization_prompt,
-                system_instruction=summary_system_message,
-            ).content
+            if self.is_langchain_compatible():
+                messages = legacy_inputs_to_messages(
+                    summarization_prompt,
+                    system_instruction=summary_system_message,
+                )
+                summary = self.llm.invoke(
+                    input=messages,
+                ).content
+            elif isinstance(self.llm, LLMInterface):
+                summary = self.llm.invoke(
+                    input=summarization_prompt,
+                    system_instruction=summary_system_message,
+                ).content
+            else:
+                raise ValueError(f"Type {type(self.llm)} of LLM is not supported.")
+
             return self.conversation_prompt(summary=summary, current_query=query_text)
         return query_text
 
+    def is_langchain_compatible(self) -> bool:
+        """Checks if the LLM is compatible with LangChain."""
+        if isinstance(self.llm, LLMInterfaceV2):
+            return True
+
+        try:
+            # langchain-core is an optional dependency
+            from langchain_core.language_models import BaseChatModel
+
+            return isinstance(self.llm, BaseChatModel)
+        except ImportError:
+            return False
+
     def _chat_summary_prompt(self, message_history: List[LLMMessage]) -> str:
         message_list = [
             f"{message['role']}: {message['content']}" for message in message_history

src/neo4j_graphrag/llm/__init__.py

@@ -16,7 +16,7 @@
 from typing import Any
 
 from .anthropic_llm import AnthropicLLM
-from .base import LLMInterface
+from .base import LLMInterface, LLMInterfaceV2
 from .cohere_llm import CohereLLM
 from .mistralai_llm import MistralAILLM
 from .ollama_llm import OllamaLLM
@@ -30,6 +30,7 @@
     "CohereLLM",
     "LLMResponse",
     "LLMInterface",
+    "LLMInterfaceV2",
     "OllamaLLM",
     "OpenAILLM",
     "VertexAILLM",