feat: refactor langchain integration with integration core by akihikokuroda · Pull Request #25 · generative-computing/mellea-contribs

akihikokuroda · 2026-02-17T18:50:32Z

Refactor of #22

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

mergify · 2026-02-17T18:51:07Z

Merge Protections

Your pull request matches the following merge protections and will not be merged until they are valid.

🟢 Enforce conventional commit

Wonderful, this rule succeeded.

Make sure that we follow https://www.conventionalcommits.org/en/v1.0.0/

title ~= ^(fix|feat|docs|style|refactor|perf|test|build|ci|chore|revert|release)(?:\(.+\))?:

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

psschwei · 2026-03-06T21:57:18Z

+        return tool_calls
+
+    @staticmethod
+    def _parse_tool_calls_from_string(content_str: str) -> list[dict[str, Any]]:


this function is very similar to the one in the base class. do we need to have a separate implementation, or can we make use of the base one?

This function is not necessary. Removing.

psschwei · 2026-03-06T21:58:12Z

+                )
+            except (ValueError, SyntaxError) as e:
+                # If parsing fails, skip this tool call
+                print(f"Warning: Failed to parse tool call arguments: {e}")


should this be a log rather than a print?

psschwei · 2026-03-06T21:59:32Z

+        # Create ChatResult with llm_output if provided
+        return ChatResult(generations=[generation], llm_output=kwargs.get("llm_output", {}))
+
+    def _extract_tool_calls(self, response: Any) -> list[dict[str, Any]]:


similar question about this one, it's also very similar to the function in the base class

I missed during refactor. Thanks!

psschwei · 2026-03-06T22:02:50Z

+        bound_model = self.__class__(
+            mellea_session=self.mellea_session,
+            model_name=self.model_name,
+            streaming=self.streaming,
+            requirements=self._requirements,
+            strategy=self._strategy,
+        )


do we need to pass kwargs too?

Suggested change

bound_model = self.__class__(

mellea_session=self.mellea_session,

model_name=self.model_name,

streaming=self.streaming,

requirements=self._requirements,

strategy=self._strategy,

)

bound_model = self.__class__(

mellea_session=self.mellea_session,

model_name=self.model_name,

streaming=self.streaming,

requirements=self._requirements,

strategy=self._strategy,

**kwargs,

)

psschwei · 2026-03-06T22:09:05Z

+        bound_model._bound_tools = tools  # type: ignore
+        bound_model._tool_choice = kwargs.get("tool_choice")  # type: ignore


we may at some point want to consider making these PrivateAttr. There may be problems down the line if someone calls model_dump() on these (which I believe both LangSmith and LangServe do), but not a high priority right now.

psschwei · 2026-03-06T22:22:59Z

+
+        Args:
+            messages: List of messages to generate from
+            stop: Stop sequences (not currently supported)


we should log if a stop sequence is passed and that it is ignored / not supported as of now

psschwei · 2026-03-06T22:26:12Z

+
+    mellea_session: Any = Field(description="The Mellea session to use for generation")
+    model_name: str = Field(default="mellea", description="Name to identify this model")
+    streaming: bool = Field(default=False, description="Whether to stream responses by default")


below (~L395) we mention that we don't support streaming, so should we drop this variable so we don't mislead?

psschwei · 2026-03-06T22:45:25Z

+        # Convert response using message converter
+        result = self.message_converter.from_mellea(response)
+
+        # Execute tool calls if present


My understanding of langchain is that the model returns a message with tool calls, but the agent executor actually runs them and sends results back. Whereas here we're running the tools and including results inside _generate. I think that may cause problems for LangGraph / LangChain agents — worth checking against langgraph's react agent implementation

Comment put this section of code.

psschwei · 2026-03-06T22:48:12Z

this is just for testing right? for a library we probably don't want to commit this

psschwei · 2026-03-06T22:50:10Z

@@ -0,0 +1,75 @@
+[build-system]
+requires = ["setuptools>=61.0", "wheel"]
+build-backend = "setuptools.build_meta"


in core we're using hatchling to build. let's standardize

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

psschwei · 2026-03-10T14:24:13Z

+This integration allows LangChain users to leverage Mellea's structured approach to generative programming while using familiar LangChain APIs. It provides full support for:
+
+- ✅ **Chat Completion**: Standard synchronous and asynchronous generation
+- ✅ **Streaming**: Real-time token-by-token generation


We probably need to update this line, as we don't have real streaming implemented in mellea yet

psschwei · 2026-03-10T14:28:27Z

+        # NOTE: Tool execution is commented out to align with standard LangChain behavior.
+        # In LangChain/LangGraph, the model returns tool calls but doesn't execute them.
+        # The agent executor handles tool execution and sends results back to the model.
+        #
+        # If you need automatic tool execution for testing or specific use cases,
+        # uncomment the code below, but be aware this deviates from standard patterns.


I would think if someone needed this functionality there should be some alternative implementation rather than having them uncomment code (which they will most likely have to go searching in their site-packages for).

The AgentExecutor has been deprecated in LangChain. For automatic tool call, it needs LangGraph. What can or should be done here?

psschwei · 2026-03-10T14:35:44Z

+    metadata: dict[str, Any]
+
+
+class MessageConverter(Protocol):


should we also add extract_last_user_message here too?

psschwei · 2026-03-10T14:37:30Z

+        # Display tool execution results if available
+        if (
+            hasattr(response, "response_metadata")
+            and "tool_execution_results" in response.response_metadata


will this still show tool execution results with the code you commented out?

I remove it.

psschwei · 2026-03-10T14:44:12Z

+    # The model only returns tool calls in the AIMessage, which the agent then executes.
+    # The _execute_tool_calls method below is for internal use and compatibility testing only.
+
+    class Config:


pydantic has changed how they define configs in v2: https://docs.pydantic.dev/latest/migration/#changes-to-config
since we're using v2 in pyproject.toml, we may want to think about updating

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

akihikokuroda · 2026-03-10T16:38:36Z

For the CI issue fix. close this and open a new one.

akihikokuroda added 6 commits February 13, 2026 15:11

feat: langchain integration

d578419

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

feat: integration base classes

4977634

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

Merge branch 'refactor' into refactor_langchain

0565e0f

feat: refactor with integration core

6e0f7a6

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

after langchain refactor

c33762f

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

Merge branch 'refactor' into refactor_langchain

62fc28c

akihikokuroda marked this pull request as draft February 17, 2026 18:50

clean ups

a418482

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

akihikokuroda marked this pull request as ready for review February 27, 2026 18:19

akihikokuroda mentioned this pull request Feb 27, 2026

feat: langchain integration #22

Closed

akihikokuroda added 3 commits March 3, 2026 09:18

add integration tests

c9c30cc

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

test enhancements

9ca58b5

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

lint error fix

5cea348

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

akihikokuroda mentioned this pull request Mar 4, 2026

feat: langchain guardrail output_parser integration #29

Merged

psschwei reviewed Mar 6, 2026

View reviewed changes

akihikokuroda added 2 commits March 6, 2026 19:58

review comments

432ea99

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

review comments

d08a8d8

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

psschwei reviewed Mar 10, 2026

View reviewed changes

review comments

8b776f5

Signed-off-by: Akihiko Kuroda <akihikokuroda2020@gmail.com>

akihikokuroda closed this Mar 10, 2026

This was referenced Mar 10, 2026

feat: langchain integration #32

Closed

feat: langchain integration #34

Merged

		bound_model._bound_tools = tools # type: ignore
		bound_model._tool_choice = kwargs.get("tool_choice") # type: ignore

Conversation

akihikokuroda commented Feb 17, 2026

Uh oh!

mergify Bot commented Feb 17, 2026

Merge Protections

🟢 Enforce conventional commit

Uh oh!

Choose a reason for hiding this comment

Uh oh!

akihikokuroda Mar 7, 2026 • edited Loading Uh oh! There was an error while loading. Please reload this page.

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

Choose a reason for hiding this comment

Uh oh!

akihikokuroda commented Mar 10, 2026

Uh oh!

Reviewers

Uh oh!

Assignees

Labels

Projects

Milestone

Development

Uh oh!

2 participants

akihikokuroda Mar 7, 2026 •

edited

Loading