Michael Limberger

Need me? Email mike@limberger.ca

This talk is about teaching a local model something new without fine-tuning. We give it documents, ask questions, and watch retrieval do the work. No training run. No cloud API required for the demo story.

The language nobody trained on

Most RAG demos cheat the test without meaning to. They use Wikipedia, popular manuals, big open datasets: pages the model has probably half-seen in training. You cannot tell whether it is reading the document or remembering it. The honest part of the demo gets lost.

So we removed the doubt. We built a real, working symbolic programming language from scratch. No English keywords. Every control structure is a symbol. No LLM (large language model, the kind of AI that writes text) has ever heard of it. If the model answers correctly, it answered from the documents we gave it. Nothing else explains it.

Three live tests

The whole session hangs on three demonstrations. Everything after that is how those moments become possible.

1. Ask a real question. Watch it cite the documentation and produce correct code.
2. Ask about a feature that does not exist yet. Watch it admit the gap instead of inventing.
3. Ask something totally off-topic. Watch it refuse cleanly.

Each one tests a different property of a healthy retrieval system. The first is retrieval quality. The second is hallucination resistance (the model inventing a fluent lie). The third is topical containment. Pass all three and the system is doing real work, not theater.

What you do not need going in

No prior AI knowledge. You do not need transformers, embeddings, fine-tuning, or vector search on the way in. We teach each word the first time it shows up. If you can use ChatGPT, you can follow this.

What you will know coming out

Why fine-tuning is the wrong tool for injecting facts, and what the research says about that. What RAG is in three sentences. How RAG-as-a-feature differs from RAG-as-infrastructure. The exact local stack: Ollama (the program that runs models on your machine), OpenWebUI (a ChatGPT-style window in the browser), and an embedding model (the searcher that turns text into numbers you can compare). And why a 730-line Perl script can outperform a Python tower of four cloud services.

Why this is not a toy

This is the architecture I shipped at the University of Toronto Faculty of Dentistry. A real production tool, real users, real policy questions. PHI (Personal Health Information, anything that identifies a patient and their health) legally cannot go to a cloud. The "never trained on" framing is a teaching device. The architecture underneath is what lets you use AI in a regulated room without leaking anything.

Is this just ChatGPT with extra steps?

No. ChatGPT is a cloud service. Everything here runs on one machine on local hardware. The model, the search, the documents, all of it.

Did we train a model for this? No, and that is the entire point. Training was the wrong tool. The next part explains why.

Can you build this yourself after? Yes, if you can install Docker (a way to run an app in a small isolated box) and a couple of terminal commands. The ten-minute setup in part 7 is a copy-paste install.

The trick is not in the model

You are about to watch a model answer correctly about a thing it has never seen. The trick is what you give it to read.