cognee：知識グラフベースのRAG構築のためのオープンソースフレームワーク、コアプロンプト学習

はじめに

Cogneeは、AIアプリケーションとAIエージェントのために設計された信頼性の高いデータレイヤーソリューションです。LLM（大規模言語モデル）コンテキストのロードと構築を行い、ナレッジグラフとベクトルストアを通じて正確で解釈可能なAIソリューションを作成するように設計されています。このフレームワークは、コスト削減、解釈可能性、ユーザーガイド可能な制御を容易にし、研究や教育用途に適しています。公式ウェブサイトでは、入門チュートリアル、概念概要、学習教材、関連研究情報を提供しています。

cogneeの最大の強みは、データを投げて自動的に処理し、ナレッジグラフを構築し、関連するトピックのグラフをつなぎ合わせることで、データ内のつながりだけでなく、関連するトピックのつながりをより明確にすることができることだ。 ラグ LLMに関しては、究極の解釈可能性を提供する。

1.データを追加し、LLMに基づいて自動的にデータを識別、処理し、ナレッジグラフに抽出し、weaviateを保存することができます。 ベクトルデータベース 2.利点は、コスト削減、データのグラフィカルな視覚化による解釈可能性、コードへの統合などによる制御可能性。

機能一覧

• ECLパイピングデータ抽出、認識、ロードを可能にし、相互接続と履歴データの検索をサポート。
• マルチデータベース対応PostgreSQL、Weaviate、Qdrant、Neo4j、Milvus、その他のデータベースをサポートします。
• 幻覚の減少パイプライン設計の最適化により、AIアプリケーションにおけるファントム現象を低減。
• 開発者フレンドリー開発者の敷居を低くするために、詳細な文書と例を提供する。
• スケーラビリティ拡張やカスタマイズが容易なモジュラー設計。

ヘルプの使用

設置プロセス

1. pipによるインストール::

   pip install cognee
   

   または、特定のデータベース・サポートをインストールする：

   pip install 'cognee[<database>]'
   

   例えば、PostgreSQLとNeo4jのサポートをインストールする：

   pip install 'cognee[postgres, neo4j]'
   

2. 詩を使ったインスタレーション::

   poetry add cognee
   

   または、特定のデータベース・サポートをインストールする：

   poetry add cognee -E <database>
   

   例えば、PostgreSQLとNeo4jのサポートをインストールする：

   poetry add cognee -E postgres -E neo4j
   

使用プロセス

1. APIキーの設定::

   import os
   os.environ["LLM_API_KEY"] = "YOUR_OPENAI_API_KEY"
   

   あるいは

   import cognee
   cognee.config.set_llm_api_key("YOUR_OPENAI_API_KEY")
   

2. .envファイルの作成.envファイルを作成し、APIキーを設定する：

   LLM_API_KEY=YOUR_OPENAI_API_KEY
   

3. LLMプロバイダーの使い分け異なるLLMプロバイダーの設定方法については、ドキュメントを参照してください。
4. 可視化結果Network を使用する場合は、Graphistry アカウントを作成し、設定します：

   cognee.config.set_graphistry_config({
   "username": "YOUR_USERNAME",
   "password": "YOUR_PASSWORD"
   })
   

主な機能

1. データ抽出複数のデータソースとフォーマットをサポートするCogneeのECLパイプラインを使用してデータを抽出します。
2. データ認識Cogneeの認知モジュールでデータを処理・分析し、幻覚を減らす。
3. データロード様々なデータベースやベクターストアをサポートしています。

注目機能 操作手順

1. 履歴データの相互接続と検索Cogneeのモジュール設計により、過去の会話、文書、音声トランスクリプションを簡単に相互接続し、検索することができます。
2. 開発者の負担軽減開発者の敷居を下げ、開発時間とコストを削減するために、詳細な文書と例を提供する。

コグニーフレームの詳細は公式サイトをご覧ください。
コグニーの理論的基礎を習得するための概要を読む
チュートリアルと学習教材で学習を始めよう

コアプロンプトコマンド

classify_content：分類されたコンテンツ

You are a classification engine and should classify content. Make sure to use one of the existing classification options nad not invent your own.
The possible classifications are:
{
"Natural Language Text": {
"type": "TEXT",
"subclass": [
"Articles, essays, and reports",
"Books and manuscripts",
"News stories and blog posts",
"Research papers and academic publications",
"Social media posts and comments",
"Website content and product descriptions",
"Personal narratives and stories"
]
},
"Structured Documents": {
"type": "TEXT",
"subclass": [
"Spreadsheets and tables",
"Forms and surveys",
"Databases and CSV files"
]
},
"Code and Scripts": {
"type": "TEXT",
"subclass": [
"Source code in various programming languages",
"Shell commands and scripts",
"Markup languages (HTML, XML)",
"Stylesheets (CSS) and configuration files (YAML, JSON, INI)"
]
},
"Conversational Data": {
"type": "TEXT",
"subclass": [
"Chat transcripts and messaging history",
"Customer service logs and interactions",
"Conversational AI training data"
]
},
"Educational Content": {
"type": "TEXT",
"subclass": [
"Textbook content and lecture notes",
"Exam questions and academic exercises",
"E-learning course materials"
]
},
"Creative Writing": {
"type": "TEXT",
"subclass": [
"Poetry and prose",
"Scripts for plays, movies, and television",
"Song lyrics"
]
},
"Technical Documentation": {
"type": "TEXT",
"subclass": [
"Manuals and user guides",
"Technical specifications and API documentation",
"Helpdesk articles and FAQs"
]
},
"Legal and Regulatory Documents": {
"type": "TEXT",
"subclass": [
"Contracts and agreements",
"Laws, regulations, and legal case documents",
"Policy documents and compliance materials"
]
},
"Medical and Scientific Texts": {
"type": "TEXT",
"subclass": [
"Clinical trial reports",
"Patient records and case notes",
"Scientific journal articles"
]
},
"Financial and Business Documents": {
"type": "TEXT",
"subclass": [
"Financial reports and statements",
"Business plans and proposals",
"Market research and analysis reports"
]
},
"Advertising and Marketing Materials": {
"type": "TEXT",
"subclass": [
"Ad copies and marketing slogans",
"Product catalogs and brochures",
"Press releases and promotional content"
]
},
"Emails and Correspondence": {
"type": "TEXT",
"subclass": [
"Professional and formal correspondence",
"Personal emails and letters"
]
},
"Metadata and Annotations": {
"type": "TEXT",
"subclass": [
"Image and video captions",
"Annotations and metadata for various media"
]
},
"Language Learning Materials": {
"type": "TEXT",
"subclass": [
"Vocabulary lists and grammar rules",
"Language exercises and quizzes"
]
},
"Audio Content": {
"type": "AUDIO",
"subclass": [
"Music tracks and albums",
"Podcasts and radio broadcasts",
"Audiobooks and audio guides",
"Recorded interviews and speeches",
"Sound effects and ambient sounds"
]
},
"Image Content": {
"type": "IMAGE",
"subclass": [
"Photographs and digital images",
"Illustrations, diagrams, and charts",
"Infographics and visual data representations",
"Artwork and paintings",
"Screenshots and graphical user interfaces"
]
},
"Video Content": {
"type": "VIDEO",
"subclass": [
"Movies and short films",
"Documentaries and educational videos",
"Video tutorials and how-to guides",
"Animated features and cartoons",
"Live event recordings and sports broadcasts"
]
},
"Multimedia Content": {
"type": "MULTIMEDIA",
"subclass": [
"Interactive web content and games",
"Virtual reality (VR) and augmented reality (AR) experiences",
"Mixed media presentations and slide decks",
"E-learning modules with integrated multimedia",
"Digital exhibitions and virtual tours"
]
},
"3D Models and CAD Content": {
"type": "3D_MODEL",
"subclass": [
"Architectural renderings and building plans",
"Product design models and prototypes",
"3D animations and character models",
"Scientific simulations and visualizations",
"Virtual objects for AR/VR environments"
]
},
"Procedural Content": {
"type": "PROCEDURAL",
"subclass": [
"Tutorials and step-by-step guides",
"Workflow and process descriptions",
"Simulation and training exercises",
"Recipes and crafting instructions"
]
}
}

generate_cog_layers：認知レイヤーを生成する

You are tasked with analyzing `{{ data_type }}` files, especially in a multilayer network context for tasks such as analysis, categorization, and feature extraction. Various layers can be incorporated to capture the depth and breadth of information contained within the {{ data_type }}.

These layers can help in understanding the content, context, and characteristics of the `{{ data_type }}`.

Your objective is to extract meaningful layers of information that will contribute to constructing a detailed multilayer network or knowledge graph.

Approach this task by considering the unique characteristics and inherent properties of the data at hand.

VERY IMPORTANT: The context you are working in is `{{ category_name }}` and the specific domain you are extracting data on is `{{ category_name }}`.

Guidelines for Layer Extraction:
Take into account: The content type, in this case, is: `{{ category_name }}`, should play a major role in how you decompose into layers.

Based on your analysis, define and describe the layers you've identified, explaining their relevance and contribution to understanding the dataset. Your independent identification of layers will enable a nuanced and multifaceted representation of the data, enhancing applications in knowledge discovery, content analysis, and information retrieval.

generate_graph_prompt: グラフ・プロンプトを生成する

You are a top-tier algorithm
designed for extracting information in structured formats to build a knowledge graph.
- **Nodes** represent entities and concepts. They're akin to Wikipedia nodes.
- **Edges** represent relationships between concepts. They're akin to Wikipedia links.
- The aim is to achieve simplicity and clarity in the
knowledge graph, making it accessible for a vast audience.
YOU ARE ONLY EXTRACTING DATA FOR COGNITIVE LAYER `{{ layer }}`
## 1. Labeling Nodes
- **Consistency**: Ensure you use basic or elementary types for node labels.
- For example, when you identify an entity representing a person,
always label it as **"Person"**.
Avoid using more specific terms like "mathematician" or "scientist".
- Include event, entity, time, or action nodes to the category.
- Classify the memory type as episodic or semantic.
- **Node IDs**: Never utilize integers as node IDs.
Node IDs should be names or human-readable identifiers found in the text.
## 2. Handling Numerical Data and Dates
- Numerical data, like age or other related information,
should be incorporated as attributes or properties of the respective nodes.
- **No Separate Nodes for Dates/Numbers**:
Do not create separate nodes for dates or numerical values.
Always attach them as attributes or properties of nodes.
- **Property Format**: Properties must be in a key-value format.
- **Quotation Marks**: Never use escaped single or double quotes within property values.
- **Naming Convention**: Use snake_case for relationship names, e.g., `acted_in`.
## 3. Coreference Resolution
- **Maintain Entity Consistency**:
When extracting entities, it's vital to ensure consistency.
If an entity, such as "John Doe", is mentioned multiple times
in the text but is referred to by different names or pronouns (e.g., "Joe", "he"),
always use the most complete identifier for that entity throughout the knowledge graph.
In this example, use "John Doe" as the entity ID.
Remember, the knowledge graph should be coherent and easily understandable,
so maintaining consistency in entity references is crucial.
## 4. Strict Compliance
Adhere to the rules strictly. Non-compliance will result in termination"""

read_query_prompt: クエリープロンプトを読む

from os import path
import logging
from cognee.root_dir import get_absolute_path

def read_query_prompt(prompt_file_name: str):
"""Read a query prompt from a file."""
try:
file_path = path.join(get_absolute_path("./infrastructure/llm/prompts"), prompt_file_name)

with open(file_path, "r", encoding = "utf-8") as file:
return file.read()
except FileNotFoundError:
logging.error(f"Error: Prompt file not found. Attempted to read: %s {file_path}")
return None
except Exception as e:
logging.error(f"An error occurred: %s {e}")
return None

render_prompt: レンダリングプロンプト

from jinja2 import Environment, FileSystemLoader, select_autoescape
from cognee.root_dir import get_absolute_path

def render_prompt(filename: str, context: dict) -> str:
"""Render a Jinja2 template asynchronously.
:param filename: The name of the template file to render.
:param context: The context to render the template with.
:return: The rendered template as a string."""

# Set the base directory relative to the cognee root directory
base_directory = get_absolute_path("./infrastructure/llm/prompts")

# Initialize the Jinja2 environment to load templates from the filesystem
env = Environment(
loader = FileSystemLoader(base_directory),
autoescape = select_autoescape(["html", "xml", "txt"])
)

# Load the template by name
template = env.get_template(filename)

# Render the template with the provided context
rendered_template = template.render(context)

return rendered_template

summarize_content：内容を要約する

You are a summarization engine and you should sumamarize content. Be brief and concise