ask.atkatana.com: Building a RAG-Powered AI from a SQLite Brain

May 2026

Every blog post, every interview answer, every career milestone and failure mode I've ever written down lives in a SQLite file on my home server. When I needed to answer a technical question about my CCoE work, I'd find myself digging through 322 entries by hand. There had to be a better way.

So I built one. ask.atkatana.com is a RAG pipeline that turns my Open Brain (OB1) knowledge base into a conversational AI — semantic search over SQLite, embeddings via local Ollama, and an LLM that answers from my own career evidence.

The Stack

```

User Question

↓

Vector Embedding (nomic-embed-text, 768-dim)

↓

ChromaDB Semantic Search → Top-K Results

↓

Context + Question → LLM (qwen3.5:9b / deepseek-r1:8b)

↓

Grounded Answer

```

Every piece is local. No API keys. No data leaves the homelab.

Open Brain (OB1) is the core — a SQLite database with 322 thoughts seeded from 12+ months of OneNote journal entries, 45+ files of corporate strategy materials, interview answers, company research, and blog content. Each thought stores text content alongside a 768-dimensional vector embedding generated by `nomic-embed-text` via Ollama.

When a question comes in, it gets embedded with the same model, ChromaDB returns the top-K most similar thoughts, and the LLM generates an answer grounded in those results. The prompt enforces citation — if the answer isn't in the retrieved context, the model says so.

Why Build This?

My CCoE work taught me that the hardest migrations are organizational, not technical. The same is true for AI adoption. Organizations need to see real, working AI that solves real problems before they can assess risk, cost, and value. A toy demo doesn't cut it.

ask.atkatana.com is that proof point. It demonstrates:

Production RAG from first principles — SQLite as vector store, local embeddings, prompt engineering for grounded output

Privacy-by-design — the entire stack runs on local hardware with no third-party API dependency

Practical infrastructure delivery — from database schema design to CI/CD deployment on a 3-host distributed homelab

The code is not the hard part. The hard part is the data: 322 curated thoughts, each one tagged, embedded, and validated. That's the same lesson I carried from the CCoE — shared services are 20% infrastructure and 80% content, standards, and organizational buy-in.

What It Means

ask.atkatana.com is a fast-hit application — a quick, focused answer to the question "can you actually ship AI?" The answer is yes, and here's the URL.

But more than that, it's a repeatable pattern. The OB1 + ChromaDB + local LLM stack is portable to any organization that needs to make its internal knowledge searchable without sending data to a third party. Healthcare, defense, financial services — the privacy argument writes itself.

The bet: organizations will want their own OB1 — a private, queryable institutional memory — before they want another SaaS chatbot. I'm already running mine.

---

*Built on a home lab, powered by local models, and owned by Andrew Katana. Try it at [ask.atkatana.com](https://ask.atkatana.com).*