Added ability to call agent with response_model and tools from openai.

Pipfile

@@ -6,45 +6,45 @@ name = "pypi"
 [packages]
 celery = { version = "==5.4.0", extras = ["redis"] }
 chromadb = "==0.5.3"
-django = "==5.0.6"
-django-allauth = "==0.63.3"
+django = "==5.1.0"
+django-allauth = "==64.1.0"
 django-appconf = "==1.0.6"
-django-celery-beat = "==2.6.0"
+django-celery-beat = "==2.7.0"
 django-csp = "==3.8"
 django-extensions = "==3.2.3"
-django-ninja = "==1.2.0"
-django-split-settings = "==1.3.1"
+django-ninja = "==1.3.0"
+django-split-settings = "==1.3.2"
 django-webpack-loader = "==3.1.0"
 esprima = "==4.0.1"
 get-docker-secret = "==2.0.0"
-instructor = "==1.3.4"
+instructor = "==1.4.0"
 ipython = "==8.26.0"
-langchain = "==0.2.6"
-langchain-chroma = "==0.1.2"
-langchain-community = "==0.2.6"
-langchain-openai = "==0.1.13"
+langchain = "==0.2.15"
+langchain-chroma = "==0.1.3"
+langchain-community = "==0.2.14"
+langchain-openai = "==0.1.23"
 langchain-text-splitters = "==0.2.2"
-litellm = "==1.41.3"
-openai = "==1.35.7"
+litellm = "==1.44.8"
+openai = "==1.42.0"
 psycopg = { version = "==3.2.1", extras = ["binary"] }
-pydantic = "==2.8.0"
+pydantic = "==2.8.2"
 python-decouple = "==3.8"
-python-gitlab = "==4.7.0"
+python-gitlab = "==4.10.0"
 redis = "==4.6.0"
 sentence-transformers = "==3.0.1"
 tantivy = "==0.22.0"
 tree-sitter = "==0.21.3"
 tree-sitter-languages = "==1.10.2"
 unidiff = "==0.7.5"
-uvicorn = "==0.30.1"
+uvicorn = "==0.30.6"
 
 [dev-packages]
-coverage = "==7.5.0"
-mypy = "==1.10.0"
-pre-commit = "==3.7.0"
+coverage = "==7.6.1"
+mypy = "==1.11.2"
+pre-commit = "==3.8.0"
 pyproject-fmt = "==1.7.0"
-pyopenssl = "==24.1.0"
-ruff = "==0.4.4"
+pyopenssl = "==24.2.1"
+ruff = "==0.6.2"
 
 [requires]
 python_version = "3.12"

daiv/automation/agents/agent.py

@@ -6,7 +6,6 @@
 from functools import cached_property
 from typing import TYPE_CHECKING
 
-import instructor
 import litellm
 from decouple import config
 from litellm import completion
@@ -26,7 +25,7 @@
 
 
 CHEAPER_MODEL = "gpt-4o-mini"
-PERFORMANT_MODEL = "gpt-4o-2024-05-13"
+PERFORMANT_MODEL = "gpt-4o-2024-08-06"
 
 
 class LlmAgent(ABC):
@@ -40,7 +39,7 @@ class LlmAgent(ABC):
     model: str = PERFORMANT_MODEL
 
     iterations: int = 0
-    max_iterations: int = 10
+    max_iterations: int = 20
 
     def __init__(
         self,
@@ -94,23 +93,22 @@ def run_iteration(self, response_model: type[BaseModel] | None = None):
 
         completion_kwargs = {"model": self.model, "messages": messages, "temperature": 0, "api_key": self.api_key}
 
-        if response_model is None:
-            response: ModelResponse = completion(
-                tools=([tool.to_schema() for tool in self.tools] if self.tools else NotGiven()), **completion_kwargs
-            )
+        response: ModelResponse = completion(
+            tools=([tool.to_schema() for tool in self.tools] if self.tools else NotGiven()),
+            response_format=response_model,
+            **completion_kwargs,
+        )
 
-            message = Message(**response.choices[0].message.model_dump())
-        else:
-            model_instance, response = instructor.from_litellm(completion).chat.completions.create_with_completion(
-                response_model=response_model, **completion_kwargs
-            )
-            message = Message(model_instance=model_instance, content=model_instance.model_dump_json(), role="assistant")
+        message = Message(**response.choices[0].message.json())
 
         self.memory.append(message)
 
         logger.debug("%s: %s", message.role, message.content)
 
-        if message.tool_calls:
+        # If the response has a model instance, avoid call the tools, it means the model has already done the job.
+        if response_model and message.content:
+            message.model_instance = response_model.model_validate_json(message.content)
+        elif message.tool_calls:
             for tool_call in message.tool_calls:
                 tool_response = self.call_tool(
                     ToolCall(
@@ -141,6 +139,9 @@ def should_continue_iteration(self, single_iteration: bool = False) -> bool:
         if single_iteration:
             return False
 
+        if self.memory[-1].model_instance is not None:
+            return False
+
         if (
             self.stop_messages
             and self.memory[-1].content

daiv/automation/coders/refactor/coder_patch.py

@@ -51,7 +51,7 @@ def invoke(self, *args, **kwargs: Unpack[MergerRequestRefactorInvoke]) -> list[F
             )
             if (
                 not (
-                    filepath_to_change := self.codebase_index.search_most_similar_filepath(
+                    filepath_to_change := self.codebase_index.search_most_similar_file(
                         kwargs["target_repo_id"], repository_file
                     )
                 )

daiv/automation/coders/refactor/prompts.py

@@ -46,41 +46,6 @@ def format_task_prompt(prompt: str) -> str:
             """
         ).format(prompt=prompt)
 
-    @staticmethod
-    def format_file_review_feedback_prompt(file_path: str, comments: list[str]) -> str:
-        """
-        Format the review feedback for the task.
-        """
-        return textwrap.dedent(
-            """\
-            ### Tasks ###
-            A developer has reviewed the file {file_path} and left comments that you need to analyze and apply changes.
-
-            ### Developer Comments ###
-            {comments}
-            """
-        ).format(file_path=file_path, comments="\n".join(comments))
-
-    @staticmethod
-    def format_diff_review_feedback_prompt(file_path: str, comments: list[tuple[str, str]]) -> str:
-        """
-        Format the review feedback for the task.
-        """
-        diff_content = ""
-        for comment in comments:
-            diff_content += f"\n{comment[0]}Hunk:\n{comment[1]}\n\n"
-
-        return textwrap.dedent(
-            """\
-            ### Tasks ###
-            A developer has reviewed the file {file_path} and left notes that you need to analyse and address one by one.
-            The review below contain a hunk of a unified diff, which includes the line number of the code where the developer left the review to locate you, and the corresponding notes to you address.
-
-            ### Developer Review ###
-            {diff_content}
-            """  # noqa: E501
-        ).format(file_path=file_path, diff_content=diff_content)
-
 
 class PatchRefactorPrompts:
     @staticmethod

daiv/automation/coders/review_addressor/coder.py

@@ -3,23 +3,33 @@
 from automation.agents.agent import LlmAgent
 from automation.agents.models import Message
 from automation.coders.base import STOP_MESSAGE, CodebaseCoder
+from automation.coders.tools import CodeActionTools, CodeInspectTools
 from automation.coders.typings import ReviewAddressorInvoke
 from codebase.base import FileChange
 
 from .models import RequestFeedback
 from .prompts import ReviewAddressorPrompts, ReviewCommentorPrompts
-from .tools import ReviewAddressorTools
 
 
-class ReviewCommentorCoder(CodebaseCoder[ReviewAddressorInvoke, RequestFeedback]):
+class ReviewCommentorCoder(CodebaseCoder[ReviewAddressorInvoke, RequestFeedback | None]):
     """
-    Coder to review the comments in the pull request.
+    Coder to review the comments left in a diff extracted from a pull request.
+
+    The coder will review the comments and ask for more information if needed.
     """
 
-    def invoke(self, *args, **kwargs: Unpack[ReviewAddressorInvoke]) -> RequestFeedback:
+    def invoke(self, *args, **kwargs: Unpack[ReviewAddressorInvoke]) -> RequestFeedback | None:
         """
         Invoke the coder to review the comments in the pull request.
         """
+        code_inspect = CodeInspectTools(
+            self.repo_client,
+            self.codebase_index,
+            self.usage,
+            repo_id=kwargs["source_repo_id"],
+            ref=kwargs["source_ref"],
+        )
+
         memory = [Message(role="system", content=ReviewCommentorPrompts.format_system(kwargs["diff"]))]
 
         for note in kwargs["notes"]:
@@ -29,27 +39,29 @@ def invoke(self, *args, **kwargs: Unpack[ReviewAddressorInvoke]) -> RequestFeedb
             else:
                 memory.append(Message(role="user", content=note.body))
 
-        self.agent = LlmAgent(memory=memory)
-        response = self.agent.run(response_model=RequestFeedback, single_iteration=True)
+        self.agent = LlmAgent(memory=memory, tools=code_inspect.get_tools())
+        response = self.agent.run(response_model=RequestFeedback)
 
         self.usage += self.agent.usage
 
         if response is None:
-            return []
+            return None
 
         return cast(RequestFeedback, response)
 
 
 class ReviewAddressorCoder(CodebaseCoder[ReviewAddressorInvoke, list[FileChange]]):
     """
-    Coder to address the review comments in the codebase.
+    Coder to address the review comments left on a pull request.
+
+    The coder will address the comments and make the necessary changes in the codebase.
     """
 
     def invoke(self, *args, **kwargs: Unpack[ReviewAddressorInvoke]) -> list[FileChange]:
         """
         Invoke the coder to address the review comments in the codebase.
         """
-        code_actions = ReviewAddressorTools(
+        code_actions = CodeActionTools(
             self.repo_client,
             self.codebase_index,
             self.usage,

daiv/automation/coders/review_addressor/models.py

@@ -9,13 +9,10 @@ class RequestFeedback(BaseModel):
     """
 
     questions: list[str] = Field(
-        default_factory=list,
         description=dedent(
             """\
             Questions for the user to answer to help you complete the task. Leave empty if there are no questions.
             """
-        ),
-    )
-    code_changes_needed: bool = Field(
-        default=False, description="Whether code changes are needed to complete the task."
+        )
     )
+    code_changes_needed: bool = Field(description="Whether code changes are needed to complete the task.")

daiv/automation/coders/review_addressor/prompts.py

@@ -14,14 +14,15 @@ def format_system(diff: str):
             ### Instructions ###
             Act as an exceptional senior software engineer that is responsible for addressing code review left on a pull request you worked on.
 
-            It's absolutely vital that you completely and correctly execute your task.
-
             The user will interact with the comments left on the code review. The unified diff has been extracted from the file where the comments were made, and shows only the specific lines of code where they were made.
 
+            It's absolutely vital that you completely and correctly execute your task.
+
             ### Guidelines ###
-            - Think out loud step-by-step before you start asking questions;
+            - Before you start asking questions:
+              * Think out loud step-by-step;
+              * Use the supplied tools to obtain more detail about the codebase and to help you think about the problem and avoid asking unnecessary questions;
             - Be straightforward on the context you need;
-            - To ask for feedback, use the provided functions;
             - Your task is completed when there's no feedback to request.
 
             ### Examples ###

daiv/automation/coders/refactor/schemas.py → daiv/automation/coders/schemas.py

@@ -23,6 +23,8 @@ class ReplaceSnippetWith(OpenAISchema):
 class CreateFile(OpenAISchema):
     """
     Use this as primary tool to create a new file with the provided content.
+
+    If the file already exists, it will raise an error. Only use this to create inexistent files.
     """
 
     file_path: str = Field(description="The file path to create.")