Cole begins by introducing Retrieval Augmented Generation (RAG) as the most popular method for giving AI agents access to knowledge bases. While acknowledging its widespread adoption and ease of implementation in no-code tools like n8n, he immediately highlights its significant limitations.

The main shortcomings identified include: RAG's inability to properly analyze trends in spreadsheets when it only retrieves fragments of tables; difficulty in finding specific documents when context is in metadata (like dates); failure to connect information across different documents; inability to "zoom out" to see entire documents; and lack of proper data analysis capabilities. These limitations form the foundation for why an improved approach is needed.

Cole introduces his solution to RAG limitations: Agentic RAG. He shows a bird's-eye view of his n8n workflow template, which appears complex but contains all the necessary components for an effective Agentic RAG system. He compares this to a previous, simpler implementation he'd created that relied solely on RAG lookups.

The key advancement in this template is the inclusion of multiple tools beyond basic RAG lookup. The agent now has several PostgreSQL tools that enable different ways of exploring the knowledge base: listing all available documents, retrieving entire file contents for specific documents, and directly querying tabular data using SQL. This gives the agent flexibility to choose the most appropriate method based on the user's question.

Cole demonstrates the capabilities of his Agentic RAG agent by testing it with several questions that highlight its ability to choose different tools based on the nature of the query. He shows how the agent handles both tabular data and text documents using the most appropriate tools for each situation.

In the first example, when asked "which month did we get the most new customers," the agent automatically writes a SQL query to analyze the revenue metrics spreadsheet rather than relying on RAG. In another example, it appropriately uses RAG to extract information about areas for improvement from a text document. Finally, Cole demonstrates how the agent can retrieve entire file contents when needed and provide source citations with clickable links to the original documents.

This section features a sponsorship segment about Unstract, an open-source tool for handling unstructured data. Cole explains that Unstract helps transform complex document types into structured data, which is particularly valuable for RAG implementations that need to work with diverse document formats.

Unstract consists of three main components: a Prompt Studio for engineering prompts that extract information from unstructured documents, Workflows for building automated extraction processes, and deployment options for creating data APIs and ETL pipelines. Cole demonstrates how easily Unstract can extract structured information from documents like receipts, making it a valuable complement to RAG systems dealing with complex document types.

Cole explains the database setup required for the Agentic RAG agent, focusing on creating three essential tables in Supabase (PostgreSQL). He walks through the SQL statements that create each table and explains their purpose in the overall system.

The three tables include: (1) a documents table that stores embeddings for RAG, metadata, and the contents of each chunk; (2) a document_metadata table for high-level information about documents including titles, URLs for source citation, and schema information for tabular data; and (3) a document_rows table that stores individual rows from spreadsheets in a JSONB format that allows SQL querying without creating separate tables for each file. Cole also provides a critical tip about using the transaction pooler connection method for Supabase to avoid connection issues.

Cole provides a detailed walkthrough of the RAG pipeline that processes documents from Google Drive into the Supabase knowledge base. The workflow begins with Google Drive triggers that detect new or updated files, then processes them through a loop to handle multiple files that might be updated simultaneously. For each file, the pipeline extracts content differently based on file type, with special handling for tabular data.

A key improvement in this pipeline is the complete deletion of existing data for a file before inserting updated content. This ensures no outdated chunks remain in the knowledge base when documents are modified. For tabular data (CSV/Excel files), the pipeline performs dual processing: it converts tables to text for RAG lookups while also storing structured data in the document_rows table for SQL querying. The pipeline also extracts schema information from spreadsheets to enable proper SQL queries later.

This section focuses on building the agent itself and configuring its tools for exploring the knowledge base. Cole explains the agent setup, which includes both webhook and chat triggers to enable API and direct interaction. The agent uses GPT-4o Mini (though Cole notes this can be upgraded) and is configured with a system prompt that guides its use of the various tools.

The four key tools provided to the agent include: (1) a RAG lookup tool that returns both content matches and metadata for source citation; (2) a document listing tool that retrieves all available documents with their IDs and metadata; (3) a file content retrieval tool that gets the full text of specific documents; and (4) a SQL query tool for analyzing tabular data. Cole explains how the agent reasons through tool selection, typically starting with RAG and progressing to other tools if needed. The SQL query tool is the most complex, requiring specific instructions on how to query the JSONB structure in the document_rows table.

Cole performs a final demonstration of the agent in action, asking it to identify employees at the company. This example effectively showcases the agent's ability to switch between different tools when one approach doesn't yield results.

The demo shows the agent first attempting a RAG lookup, determining that it didn't find what it needed, then listing available documents to identify which might contain employee information. After identifying the product team meeting minutes as a likely source, it retrieves the full content of that document to answer the question. Cole notes that this spontaneous example worked perfectly to demonstrate the key advantage of Agentic RAG: the ability to seamlessly switch between different knowledge retrieval methods when one approach fails.

In the closing segment, Cole summarizes the value of the template he's created and hints at future content. He emphasizes that while the template is advanced, it provides a solid starting point for implementing Agentic RAG in n8n, and encourages viewers to ask questions in the comments if they encounter issues while building out the workflow.

Cole also announces an upcoming video featuring a completely local version of the Agentic RAG agent built with the LocalAI package, suggesting that those interested in self-hosted solutions should stay tuned. He concludes by asking viewers to like and subscribe if they found the content valuable and are looking forward to more content on AI agents and n8n.

The Agentic RAG approach demonstrated in this video represents a significant evolution in how AI agents interact with knowledge bases. By moving beyond simple vector lookups to a multi-tool approach, Cole's template enables more reliable, versatile, and intelligent information retrieval that overcomes many limitations of traditional RAG implementations.

The true power of this approach lies in its adaptability. When faced with tabular data, the agent can perform precise SQL queries instead of relying on fragmented text chunks. When context from an entire document is needed, it can retrieve and analyze complete files. When documents need to be selected based on metadata, it can browse and reason about the available options. This flexibility results in a much more capable knowledge assistant that can handle a diverse range of queries with greater accuracy.

So what? For developers and no-code builders working with AI agents, this template provides not just a ready-to-use solution but a framework for thinking about knowledge retrieval in a more sophisticated way. Rather than forcing all queries through a single retrieval method, the Agentic RAG approach encourages us to give our AI agents a toolbox of complementary methods and the reasoning ability to select the right tool for each job. This paradigm shift—from rigid, single-path knowledge retrieval to flexible, multi-tool exploration—represents the future of building truly helpful AI assistants that can work effectively with diverse knowledge bases.