What is the elegant way to build a proper knowledge base?
It depends on your purpose – you don’t need a knowledge base for everything, but it’s useful for letting an AI answer business‑specific questions that would otherwise require a database query. For example, I would feed all publicly available information about me, my company (Left Click Software), and our major case studies into the system. I’d host the chatbot on a platform like Botpress or NLU and connect it directly to Claude via API. The system prompt would contain a simple template with a few conversation routes to keep answers constrained. All the data would be embedded in a vector store such as Pinecone, and Claude would query Pinecone for the top five to ten matches before responding. This is standard RAG: each incoming message triggers the system prompt, a Pinecone similarity search, injection of the results into the prompt, and then the answer. The key is to be specific – you can’t feed everything – and to consider security, since exposing internal data through a public model can leak sensitive information.
I haven't done it because I find there's very little business utility—these tools tend to work only about 80% of the time and the response quality is pretty poor, so most of this functionality is below the level where businesses would pay for it. My channel is all about businesses, strategies, and systems that businesses will actually pay money for. That said, I could definitely show the simplest, straightforward way to get up and running with this on your own, whether you want it as a hobby project or see an opportunity to monetize it. Here’s how retrieval‑augmented generation works: you feed a bunch of text into a database; the database indexes the text—turning it into numbers, essentially a high‑dimensional vector representation. When you ask a question, the system searches that database for snippets of text similar to your query (like Googling over your own data), retrieves those matches, and inserts them into the prompt for a large language model. The model then answers based on both the retrieved data and its own built‑in intelligence. To illustrate, I drew a quick diagram: RAG stands for retrieval‑augmented generation. Retrieval is just Googling over your own dataset—user inputs a query, the query searches the database, we find however many matches we want, we insert those matches into the LLM’s prompt, and the LLM answers the question. There are many ways to implement this, but the simplest way I’ll cover is using the OpenAI Assistants playground. In the playground you define the assistant’s instructions, then under Tools you enable File Search, upload a text file (I used one of my blog posts, stripped of formatting), and the system automatically creates a vector store and handles the embedding and retrieval for you. As a demo, I asked the assistant, “What are three components to a great high‑ROI automation?” It returned: a clear deliverable, it solves a hot‑button pain point, and you can template most of it. I also showed how to define a “pressing problem” as anything that impacts your ability to generate revenue, and the assistant gave a relevant answer. The result is your own RAG‑enabled chatbot. There are various services—like Chatbase, Botpress, or others—that let you run assistants through their back ends, or they may have their own RAG implementations to charge you for. You can embed this onto your website or service in many ways. I don’t sell it to businesses because I don’t find it very valuable for them, but chatbots are growing in popularity and functionality, so there’s definitely potential there. Hopefully that answers your question—thanks for asking it!
Here’s how I’d build a fully autonomous reply bot for a cold‑email campaign that sends 10 000+ messages a day. I’d start with a cheap inference platform such as N8N and create an AI agent inside it. The agent would have access to a knowledge base about the company and a library of proven reply templates collected from the past few months.
When a new inbound email arrives, a workflow would trigger that (1) pulls the entire conversation history with that prospect, (2) checks the knowledge base to confirm the company context, and (3) scans the template library to find the closest match for the desired outcome (e.g., booking a meeting). The selected template would be populated with any dynamic details, and the AI would generate the final reply while staying as close to the template as possible to avoid sounding overly AI‑generated.
The idea is to give the AI minimal freedom—just enough to fill in variables—so you retain control over tone and quality. I originally built this in N8N, but once you scale to 100 000 emails per day you can replace the visual builder with lightweight Python scripts running on a serverless platform like Modal, listening to webhook events from your email service.
When selling chatbots, emphasize the outcome they deliver rather than the fact that they chat in a human‑like way. Position yourself as selling speed‑to‑lead systems that, on average, double a client’s top of funnel or generate about $15,000 for businesses earning $50,000‑$100,000 in the digital‑agency space. This avoids the commodity race to the bottom where chatbots sell for $10/month. For learning chatbots today, it’s all about prompting. You don’t need RAG or retrieval‑augmented generation. Simply take a model like GPT‑4.1 Mini (million‑token context), load the client’s entire knowledge base—articles, PDFs, OCR‑converted text—into the prompt, compress if necessary by telling the AI to remove whitespace and superfluous words, and you can fit roughly thirty books of content in a single prompt.
Yeah, absolutely — it's just a matter of time. We're already getting to the point where you can have really good, seemingly customized conversations with them, just using GPT-4.5 (the new model that just dropped about a week before this recording, and it's very good). I have no doubt that the next generation of models will let you simply say "hey, this is the business, here's some stuff they do, do it" — it'll automatically pick up the tone of voice, automatically pick up what you want it to do, and even ask you a clarifying question when you initialize it, like "do you want me to book appointments? What do you want me to do?" A lot of this is going to get done for you. When that happens, your point of leverage changes — it stops being about actually installing and implementing a chatbot with a bunch of agent features doing function calling or tool calling, and becomes about knowing where the tools are and the simplest, easiest way to deploy them for the customer. So yeah, we'll get there pretty soon — probably not for at least another year, maybe six months to a year — so there's still a lot of value in knowing how to do it manually right now.