Structured code summarization (#375)

* feat: turn summarize_code into generator

* feat: extract run_code_graph_pipeline, update the pipeline

* feat: minimal code graph example

* refactor: update argument

* refactor: move run_code_graph_pipeline to cognify/code_graph_pipeline

* refactor: indentation and whitespace nits

* refactor: add deprecated use comments and warnings

* Structured code summarization

* add missing prompt file

* Remove summarization_model argument from summarize_code and fix typehinting

* minor refactors

---------

Co-authored-by: lxobr <[email protected]>
Co-authored-by: Vasilije <[email protected]>
Co-authored-by: Igor Ilic <[email protected]>
Co-authored-by: Boris <[email protected]>

cognee/api/v1/cognify/code_graph_pipeline.py

@@ -7,22 +7,27 @@
 from pathlib import Path
 from typing import Union
 
-from cognee.shared.SourceCodeGraph import SourceCodeGraph
-from cognee.shared.data_models import SummarizedContent
-from cognee.shared.utils import send_telemetry
-from cognee.modules.data.models import Dataset, Data
-from cognee.modules.data.methods.get_dataset_data import get_dataset_data
 from cognee.modules.data.methods import get_datasets, get_datasets_by_name
-from cognee.modules.pipelines.tasks.Task import Task
+from cognee.modules.data.methods.get_dataset_data import get_dataset_data
+from cognee.modules.data.models import Data, Dataset
 from cognee.modules.pipelines import run_tasks
-from cognee.modules.users.models import User
-from cognee.modules.users.methods import get_default_user
 from cognee.modules.pipelines.models import PipelineRunStatus
-from cognee.modules.pipelines.operations.get_pipeline_status import get_pipeline_status
-from cognee.modules.pipelines.operations.log_pipeline_status import log_pipeline_status
-from cognee.tasks.documents import classify_documents, check_permissions_on_documents, extract_chunks_from_documents
+from cognee.modules.pipelines.operations.get_pipeline_status import \
+    get_pipeline_status
+from cognee.modules.pipelines.operations.log_pipeline_status import \
+    log_pipeline_status
+from cognee.modules.pipelines.tasks.Task import Task
+from cognee.modules.users.methods import get_default_user
+from cognee.modules.users.models import User
+from cognee.shared.SourceCodeGraph import SourceCodeGraph
+from cognee.shared.utils import send_telemetry
+from cognee.tasks.documents import (check_permissions_on_documents,
+                                    classify_documents,
+                                    extract_chunks_from_documents)
 from cognee.tasks.graph import extract_graph_from_code
-from cognee.tasks.repo_processor import get_repo_file_dependencies, enrich_dependency_graph, expand_dependency_graph
+from cognee.tasks.repo_processor import (enrich_dependency_graph,
+                                         expand_dependency_graph,
+                                         get_repo_file_dependencies)
 from cognee.tasks.storage import add_data_points
 from cognee.tasks.summarization import summarize_code
 
@@ -134,7 +139,7 @@ async def run_code_graph_pipeline(repo_path):
         Task(get_repo_file_dependencies),
         Task(enrich_dependency_graph, task_config={"batch_size": 50}),
         Task(expand_dependency_graph, task_config={"batch_size": 50}),
-        Task(summarize_code, summarization_model=SummarizedContent, task_config={"batch_size": 50}),
+        Task(summarize_code, task_config={"batch_size": 50}),
         Task(add_data_points, task_config={"batch_size": 50}),
     ]
 

cognee/infrastructure/llm/prompts/summarize_code.txt

@@ -0,0 +1,10 @@
+You are an expert Python programmer and technical writer. Your task is to summarize the given Python code snippet or file.
+The code may contain multiple imports, classes, functions, constants and logic. Provide a clear, structured explanation of its components 
+and their relationships.
+
+Instructions:
+Provide an overview: Start with a high-level summary of what the code does as a whole.
+Break it down: Summarize each class and function individually, explaining their purpose and how they interact.
+Describe the workflow: Outline how the classes and functions work together. Mention any control flow (e.g., main functions, entry points, loops).
+Key features: Highlight important elements like arguments, return values, or unique logic.
+Maintain clarity: Write in plain English for someone familiar with Python but unfamiliar with this code.

cognee/modules/data/extraction/extract_summary.py

@@ -1,7 +1,11 @@
 from typing import Type
+
 from pydantic import BaseModel
-from cognee.infrastructure.llm.prompts import read_query_prompt
+
 from cognee.infrastructure.llm.get_llm_client import get_llm_client
+from cognee.infrastructure.llm.prompts import read_query_prompt
+from cognee.shared.data_models import SummarizedCode
+
 
 async def extract_summary(content: str, response_model: Type[BaseModel]):
     llm_client = get_llm_client()
@@ -11,3 +15,7 @@ async def extract_summary(content: str, response_model: Type[BaseModel]):
     llm_output = await llm_client.acreate_structured_output(content, system_prompt, response_model)
 
     return llm_output
+
+async def extract_code_summary(content: str):
+
+    return await extract_summary(content, response_model=SummarizedCode)

cognee/shared/data_models.py

@@ -1,9 +1,11 @@
 """Data models for the cognitive architecture."""
 
 from enum import Enum, auto
-from typing import Optional, List, Union, Dict, Any
+from typing import Any, Dict, List, Optional, Union
+
 from pydantic import BaseModel, Field
 
+
 class Node(BaseModel):
     """Node in a knowledge graph."""
     id: str
@@ -194,6 +196,29 @@ class SummarizedContent(BaseModel):
     summary: str
     description: str
 
+class SummarizedFunction(BaseModel):
+    name: str
+    description: str
+    inputs: Optional[List[str]] = None
+    outputs: Optional[List[str]] = None
+    decorators: Optional[List[str]] = None
+
+class SummarizedClass(BaseModel):
+    name: str
+    description: str
+    methods: Optional[List[SummarizedFunction]] = None
+    decorators: Optional[List[str]] = None
+
+class SummarizedCode(BaseModel):
+    file_name: str
+    high_level_summary: str
+    key_features: List[str]
+    imports: List[str] = []
+    constants: List[str] = []
+    classes: List[SummarizedClass] = []
+    functions: List[SummarizedFunction] = []
+    workflow_description: Optional[str] = None
+
 
 class GraphDBType(Enum):
     NETWORKX = auto()

cognee/tasks/summarization/summarize_code.py

@@ -1,39 +1,36 @@
 import asyncio
+from typing import AsyncGenerator, Union
 from uuid import uuid5
 from typing import Type
 
-from pydantic import BaseModel
-
 from cognee.infrastructure.engine import DataPoint
-from cognee.modules.data.extraction.extract_summary import extract_summary
-from cognee.shared.CodeGraphEntities import CodeFile
+from cognee.modules.data.extraction.extract_summary import extract_code_summary
 from .models import CodeSummary
 
 
 async def summarize_code(
     code_graph_nodes: list[DataPoint],
-    summarization_model: Type[BaseModel],
-) -> list[DataPoint]:
+) -> AsyncGenerator[Union[DataPoint, CodeSummary], None]:
     if len(code_graph_nodes) == 0:
         return
 
-    code_files_data_points = [file for file in code_graph_nodes if isinstance(file, CodeFile)]
+    code_data_points = [file for file in code_graph_nodes if hasattr(file, "source_code")]
 
     file_summaries = await asyncio.gather(
-        *[extract_summary(file.source_code, summarization_model) for file in code_files_data_points]
+        *[extract_code_summary(file.source_code) for file in code_data_points]
     )
 
     file_summaries_map = {
-        code_file_data_point.extracted_id: file_summary.summary
-        for code_file_data_point, file_summary in zip(code_files_data_points, file_summaries)
+        code_data_point.extracted_id: str(file_summary)
+        for code_data_point, file_summary in zip(code_data_points, file_summaries)
     }
 
     for node in code_graph_nodes:
         if not isinstance(node, DataPoint):
             continue
         yield node
 
-        if not isinstance(node, CodeFile):
+        if not hasattr(node, "source_code"):
             continue
 
         yield CodeSummary(