Workflow Pre-requisites

Step 1: Supabase Setup

First, replace the keys in the "Save the embedding in DB" & "Search Embeddings" nodes with your new Supabase keys. After that, run the following code snippets in your Supabase SQL editor:

1. Create the table to store chunks and embeddings:

   CREATE TABLE public."RAG"
   (
       id bigserial PRIMARY KEY,
       chunk text NULL,
       embeddings vector(1024) NULL
   )
   TABLESPACE pg_default;
   

2. Create a function to match embeddings:

   DROP FUNCTION IF EXISTS public.matchembeddings1(integer, vector);
   
   CREATE OR REPLACE FUNCTION public.matchembeddings1(
       match_count integer,
       query_embedding vector
   )
   RETURNS TABLE (
       chunk text,
       similarity float
   )
   LANGUAGE plpgsql
   AS $$
   BEGIN
       RETURN QUERY
       SELECT
           R.chunk,
           1 - (R.embeddings <=> query_embedding) AS similarity
       FROM public."RAG" AS R
       ORDER BY R.embeddings <=> query_embedding
       LIMIT match_count;
   END;
   $$;
   

Step 2: Create Jotform with these fields

1. Your full name
2. email address
3. Upload PDF Document [field where you upload the knowledgebase in PDF]

Step 3: Get Together AI API Key

Get a Together AI API key and paste it into the "Embedding Uploaded document" node and the "Embed User Message" node.

Here is a detailed, node-by-node explanation of the n8n workflow, which is divided into two main parts.

Part 1: Ingesting Knowledge from a PDF

This first sequence of nodes runs when you submit a PDF through a Jotform. Its purpose is to read the document, process its content, and save it in a specialized database for the AI to use later.

1. JotForm Trigger

   • Type: Trigger
   • What it does: This node starts the entire workflow. It's configured to listen for new submissions on a specific Jotform. When someone uploads a file and submits the form, this node activates and passes the submission data to the next step.

2. Grab New knowledgebase

   • Type: HTTP Request
   • What it does: The initial trigger from Jotform only contains basic information. This node makes a follow-up call to the Jotform API using the submissionID to get the complete details of that submission, including the specific link to the uploaded file.

3. Grab the uploaded knowledgebase file link

   • Type: HTTP Request
   • What it does: Using the file link obtained from the previous node, this step downloads the actual PDF file. It's set to receive the response as a file, not as text.

4. Extract Text from PDF File

   • Type: Extract From File
   • What it does: This utility node takes the binary PDF file downloaded in the previous step and extracts all the readable text content from it. The output is a single block of plain text.

5. Splitting into Chunks

   • Type: Code
   • What it does: This node runs a small JavaScript snippet. It takes the large block of text from the PDF and chops it into smaller, more manageable pieces, or "chunks," each of a predefined length. This is critical because AI models work more effectively with smaller, focused pieces of text.

6. Embedding Uploaded document

   • Type: HTTP Request
   • What it does: This is a key AI step. It sends each individual text chunk to an embeddings API. A specified AI model converts the semantic meaning of the chunk into a numerical list called an embedding or vector. This vector is like a mathematical fingerprint of the text's meaning.

7. Save the embedding in DB

   • Type: Supabase
   • What it does: This node connects to your Supabase database. For every chunk, it creates a new row in a specified table and stores two important pieces of information: the original text chunk and its corresponding numerical embedding (its "fingerprint") from the previous step.

Part 2: Answering Questions via Chat

This second sequence starts when a user sends a message. It uses the knowledge stored in the database to find relevant information and generate an intelligent answer.

1. When chat message received

   • Type: Chat Trigger
   • What it does: This node starts the second part of the workflow. It listens for any incoming message from a user in a connected chat application.

2. Embend User Message

   • Type: HTTP Request
   • What it does: This node takes the user's question and sends it to the exact same embeddings API and model used in Part 1. This converts the question's meaning into the same kind of numerical vector or "fingerprint."

3. Search Embeddings

   • Type: HTTP Request
   • What it does: This is the "retrieval" step. It calls a custom database function in Supabase. It sends the question's embedding to this function and asks it to search the knowledge base table to find a specified number of top text chunks whose embeddings are mathematically most similar to the question's embedding.

4. Aggregate

   • Type: Aggregate
   • What it does: The search from the previous step returns multiple separate items. This utility node simply bundles those items into a single, combined piece of data. This makes it easier to feed all the context into the final AI model at once.

5. AI Agent & Google Gemini Chat Model

   • Type: LangChain Agent & AI Model
   • What it does: This is the "generation" step where the final answer is created.
     • The AI Agent node is given a detailed set of instructions (a prompt).
     • The prompt tells the Google Gemini Chat Model to act as a professional support agent.
     • Crucially, it provides the AI with the user's original question and the aggregated text chunks from the Aggregate node as its only source of truth.
     • It then instructs the AI to formulate an answer based only on that provided context, format it for a specific chat style, and to say "I don't know" if the answer cannot be found in the chunks. This prevents the AI from making things up.