I Built the ULTIMATE n8n RAG AI Agent Template

Key Concepts

Retrieval Augmented Generation (RAG), Agentic RAG, Knowledge Base, n8n, Superbase, Vector Embeddings, SQL Queries, Data Analysis, AI Agents, Tool Use, Prompt Engineering, Data Pipelines, Unstructured Data, Structured Data, Document Chunking, Metadata.

Agentic RAG: Overcoming Limitations of Traditional RAG

The video addresses the limitations of traditional Retrieval Augmented Generation (RAG) and introduces Agentic RAG as a solution. Traditional RAG often struggles with:

• Missing Key Context: RAG's lookup mechanism can miss crucial context and related information, especially when dealing with large documents or datasets.
  • Example: Analyzing trends in a spreadsheet where RAG only pulls a fraction of the table.
• Incorrect Document Retrieval: RAG can fail to retrieve the correct document, even when the relevant information (e.g., date) is clearly present.
  • Example: Summarizing meeting notes but RAG pulls notes from the wrong date.
• Lack of Cross-Document Connection: RAG often struggles to connect different documents to provide broader context.
• Inability to Zoom Out: RAG cannot effectively analyze entire documents or sets of documents unless the context is small enough.
• No Data Analysis Capabilities: RAG lacks the ability to perform proper data analysis, such as calculating sums or trends.

Agentic RAG overcomes these limitations by:

• Enabling Reasoning: Giving AI agents the ability to reason about how they explore the knowledge base.
• Multiple Tools: Providing agents with multiple tools to explore the knowledge base instead of just a single RAG lookup.
• Query Improvement: Allowing agents to improve RAG lookup queries.
• Tool Selection: Enabling agents to choose different tools to answer different user questions.

Definition of Agentic RAG: Agentic RAG empowers agents to reason about how they explore a knowledge base, offering multiple tools and the ability to refine queries and select appropriate tools based on the user's question.

Agentic RAG Agent Architecture in n8n

The video presents a detailed walkthrough of an Agentic RAG agent built in n8n, a no-code workflow automation platform. The architecture consists of the following key components:

1. RAG Pipeline

The RAG pipeline is responsible for ingesting and processing documents from various sources (e.g., Google Drive) and storing them in a Superbase knowledge base. The pipeline includes the following steps:

• Document Trigger: Monitors a specified folder in Google Drive for new or updated files.
  • Details: Uses Google Drive trigger to poll every minute for new files. Handles both file creation and updates.
• File Type Handling: Determines the file type and uses appropriate extraction methods.
  • Details: Uses a switch node to route files based on type (CSV, Google Doc, etc.).
• Data Cleaning: Clears out old data for the file in Superbase to ensure data consistency.
  • Details: Deletes existing document rows and data rows in Superbase based on file ID.
• Metadata Management: Inserts or updates metadata for the document in Superbase.
  • Details: Stores high-level information like title, URL, and schema (for tabular data).
• Content Extraction: Extracts the content from the document.
  • Details: Uses dedicated nodes for different file types (e.g., "Extract Text Document" for PDFs, "Extract CSV" for CSV files).
• Data Transformation: Transforms the extracted content into a suitable format for RAG and data analysis.
  • Details: For CSV files, converts data into rows for SQL querying and aggregates data into a text document for RAG.
• Vector Embedding: Creates vector embeddings for the document content using an embedding model (e.g., OpenAI's text-embedding-3-small).
  • Details: Uses a default data loader for chunking documents and defining metadata.
• Superbase Integration: Inserts the document content and embeddings into the Superbase vector store.
  • Details: Stores embeddings, metadata (file ID, file title), and content in the documents table.

2. AI Agent

The AI agent is responsible for interacting with the user, understanding their questions, and using the available tools to retrieve and analyze information from the knowledge base. The agent includes the following components:

• Triggers: Webhook and chat triggers to initiate the agent.
• Agent Node: Uses a system prompt to define the agent's behavior and available tools.
  • Details: System prompt instructs the agent to use RAG first and then other tools if necessary.
• Tools: A set of tools that the agent can use to explore the knowledge base.
  • RAG Lookup: Performs a vector search in the Superbase vector store to retrieve relevant document chunks.
    • Details: Includes metadata (file ID, file title) in the results for source citation.
  • List Documents: Retrieves a list of all documents in the knowledge base.
    • Details: Queries the document_metadata table to get document titles, IDs, and schemas.
  • Get File Contents: Retrieves the full content of a specific document.
    • Details: Queries the documents table to retrieve all chunks for a given file ID.
  • Query Excel/CSV: Executes SQL queries against CSV and Excel files stored in Superbase.
    • Details: Uses the document_rows table and the row_data JSONB column to query tabular data.

Tool Use and Reasoning

The video demonstrates how the Agentic RAG agent uses different tools to answer user questions. The agent can:

• Use RAG for general information retrieval.
  • Example: Answering "What are areas we could do better with?" by retrieving relevant chunks from a customer feedback survey.
• Use the "List Documents" tool to identify relevant files.
  • Example: Listing available documents to find meeting notes for a specific date.
• Use the "Get File Contents" tool to retrieve the full content of a specific file.
  • Example: Retrieving the action items from a product team meeting minutes document.
• Use the "Query Excel/CSV" tool to perform data analysis on tabular data.
  • Example: Determining which month had the most new customers by querying a revenue metrics spreadsheet.

The agent's ability to reason about which tool to use based on the user's question is a key advantage of Agentic RAG.

Unstructured Data Processing with Unraq

The video highlights the importance of handling unstructured data for AI agents and introduces Unraq, an open-source no-code LLM platform, as a solution. Unraq helps to:

• Turn unstructured documents into structured data.
• Extract information from PDFs, images, and other complex document formats.
• Create APIs and ETL pipelines for data processing.

Unraq consists of three main parts:

• Prompt Studio: Allows users to engineer prompts for extracting information from unstructured documents.
• Workflows: Enables users to build flows for automatically extracting information from documents.
• API Deployments and ETL Pipelines: Allows users to deploy workflows as data APIs and ETL pipelines.

Setting up Superbase Database

The video outlines the steps to set up the Superbase database for the Agentic RAG agent:

1. Create the documents table: Stores embeddings, metadata, and content for each document chunk.
   • SQL Query: (Example from the video)

     CREATE TABLE documents (
         id UUID PRIMARY KEY DEFAULT gen_random_uuid(),
         content TEXT,
         metadata JSONB,
         embedding VECTOR(1536)
     );
     

2. Create the document_metadata table: Stores high-level information for each document, such as title, URL, and schema.
3. Create the document_rows table: Stores data rows for CSV and Excel files in JSONB format.

Key Takeaways

• Agentic RAG overcomes the limitations of traditional RAG by enabling agents to reason about how they explore a knowledge base and providing them with multiple tools.
• The n8n workflow provides a practical example of how to build an Agentic RAG agent.
• Unraq is a valuable tool for handling unstructured data and extracting information from complex document formats.
• Properly setting up the Superbase database is crucial for storing and retrieving information for the Agentic RAG agent.
• Prompt engineering and tool selection are key to the success of Agentic RAG.

Conclusion

The video provides a comprehensive overview of Agentic RAG and its advantages over traditional RAG. It offers a detailed walkthrough of an n8n workflow for building an Agentic RAG agent, including the RAG pipeline, AI agent, and Superbase database setup. The video also highlights the importance of handling unstructured data and introduces Unraq as a solution. The key takeaway is that Agentic RAG empowers AI agents to explore knowledge bases more effectively and answer user questions more accurately by providing them with multiple tools and the ability to reason about how to use them.