Yesterday I managed to get the full document into the vector database. Chunking, embeddings, the table setup in Supabase, everything worked. Yessss 😁 So today I want to take the next logical step. I want to actually talk to the vectorized data. Chat with it. Ask questions. See what kind of answers I get and how good the quality is, as far as I can judge at this point. This is the exciting part for me. Yesterday was all about structure and setup. Today is about seeing whether the whole pipeline actually produces meaningful results. If the retrieval works. If the embeddings make sense. If the answers stay close to the source. I know the answers won’t be perfect. But that is exactly why I’m doing this challenge. Now that the data is in the vector store, the next step is to create the retrieval flow in n8n, send a query, look at the matches, and see how the LLM responds with context. Let’s see how good or bad the first results will be. Time to experiment again.

Today I stayed in n8n again. As I mentioned earlier, I need to understand the basics here before I switch to Claude Code with a full RAG setup. So my goal for today was simple.I wanted to load a PDF from Google Drive into a vector database in Supabase. I created a new project in Supabase. From what I can see, you can have two active projects before the paid version starts, which is perfect for experimenting. I added the API key and URL to the Supabase node in n8n. Next I needed a table. So I followed the docs under “quickstart for setting up your vector store,” copied the SQL snippet and pasted it into the SQL editor. Unfortunately I got an error. I copied the code and the error message into Google Gemini, added a short description of what I wanted to do and it immediately generated a corrected version. I ran it again and the table was created. Really impressive. After that I added the data loader in n8n for chunking the document. I kept the default settings for now. Chunk size 1000 and overlap 200. For embeddings I connected the OpenAI embeddings model. Apart from a small code issue, which Gemini also helped me fix very quickly, everything worked. Then I ran the full flow. And it was amazing to see how the document was split into 77 items and then stored in the vector database. One by one in the table. A really successful day. The data is now in the vector store and I can honestly say that I learned something valuable again today.

Today I worked on integrating Supabase as an MCP inside Claude Code. Throughout November, I’ve spent at least one hour every day diving into Claude Code and exploring all the possibilities. But honestly? I still have no idea how to actually get a RAG with Claude Code and Supabase to work. I watched a video about it, but it didn’t really click for me. I can set up the MCP connection but what happens after that? How do I get from there to a proper RAG setup? Today I’m just missing the imagination for how everything connects. And every time I’m surprised by how fast one hour flies by. I fully agree with Mark: long-term, Claude Code is absolutely the right direction. And I will definitely go that way. But I need much more understanding first. A lot more, to be honest. So for today, I’ve decided to continue down the n8n path. I’m going to gather all the RAG automations I can find across different communities and try to rebuild them step by step. I hope that in a few days I’ll have enough understanding to take the Claude Code + Supabase MCP route with confidence. What do you think?Any good suggestions? Or maybe you’ve already built a cool RAG setup with Claude Code that you’d like to share here? Also: Do you have any interesting YouTube material I should study?

Today I thought I would finally build something simple.The goal was easy. Just send a basic message from n8n to Pinecone, talk to the assistant and get a response from the patent document. That was the plan.And of course, it didn’t go that way. I set up the manual trigger, added the AI Agent node, connected the model, added the HTTP request to Pinecone and expected everything to run smoothly. But no.The first thing I got was an error from the AI Agent node: This model is not supported in version 2.2 of the Agent node. I had no idea why. I tried different things, tested several setups, switched models and still no luck. In the end I used OpenRouter.And suddenly it worked. I still don’t really know why this fixed it. It just cost a lot of time and nerves. The next problem was the system prompt, it took longer than expected because the AI Agent refused to call Pinecone. It kept acting weird and simply didn’t follow the instructions. But eventually I got it to work.I can now chat with the document inside Pinecone through the assistant. This took way more than one hour. More as i had planned and the system prompt still gives me trouble. Sometimes it works, sometimes it doesn’t, and I can’t explain why yet. But at least something is working now. But my question,.... Why? I will continue you can be sure

Today I looked at the issue from earlier again to really understand what was going on. Someone in the community gave me a simple and very helpful tip.Just extract the text from the PDF and convert it into a .md file.For now this is the easiest solution. I tried it and it worked. I pulled the text out of the PDF, saved it as a markdown file, uploaded it and Pinecone finally accepted it. The text was processed without any problems. I still don’t fully understand why the original PDF didn’t work.But I will figure it out sooner or later. In the meantime I started thinking about how to automate this.Maybe I can use an OCR step to extract text from any PDF and then pass that text to Pinecone.Mistral has OCR. Do you know any other good options I should look at? I also want to test Supabase later to compare how it works with a simple RAG setup. Now the next step is to send everything into n8n and see what happens. Let’s keep going my next fight