How we set ours up · 27 Oct 2026

654,718 chunks, searched on one device

You commented OWN on the Dessa reel. This is how we set ours up, and how you can install the first two parts today.

Four parts run on one device, with no AI bill. Two of them search our notes and records, cut into short pieces we call chunks.

1Can you use Dessa?

Block released Dessa on 23 October. It listens and talks at the same time.

Block says it runs on a laptop: a MacBook Pro M4 Max or better.

But you can't use it yet. It's a research preview, and only a small group of AI safety researchers can try it. Block has a waitlist (block.xyz/dessa).

So run an open model on your own machine instead. We run Ollama and three open models that anyone can download today.

2What we run

Our search covers 654,718 chunks of our notes and records. That's the count on 27 October. The reel said almost six hundred thousand; it has grown since.

Our search refreshes four times a day. It only re-embeds (turns into numbers again) the chunks that changed.

3Start in three steps

1. Install Ollama. On a Mac or Linux, paste the first line below into Terminal. On Windows, get the app from ollama.com/download, then type steps 2 and 3 in a terminal window.
2. Get the model that makes your notes searchable. Run ollama pull nomic-embed-text.
3. Get a model to talk to. Run ollama run qwen3:8b (5.2GB), then type a question.

curl -fsSL https://ollama.com/install.sh | sh
ollama pull nomic-embed-text
ollama run qwen3:8b

You now have the model our lead sorter tries first, on your own machine.

4Point it at your notes

1. Cut your files into chunks. Ours are about 1,200 characters, split at headings, with 200 characters of overlap.
2. Turn each chunk into numbers with nomic-embed-text. Ollama does this on your own machine.
3. Build a keyword index too. Ours uses BM25.
4. Ask both and merge the answers. We merge the two ranked lists into one, by reciprocal-rank fusion.
5. Only redo what changed. Keep a hash (a short code) of each chunk, and skip the ones that haven't changed.

Our search only talks to Ollama on the same computer, so the search never sends our notes anywhere.

5Turn speech into text

Whisper transcribes our videos. Install it with Homebrew: brew install openai-whisper. It's open, under the MIT licence.

We run whisper's tiny model by default, up to medium.

6The honest limits

• An 8B model isn't a frontier model. We use it for sorting messages.
• You need the disk space. The two models above take about 5.5GB. The computer also needs enough memory to run the model. Check the model's page on ollama.com.
• Nothing leaves the machine as long as you keep it that way. Our search only calls its own computer. Add a cloud service and that changes.

7Hand this to your agent

If you have a developer or an AI agent that can work on your computer, give it this.

Set up a local AI search over my notes on this computer. Ask me before installing anything.

1. Install Ollama from ollama.com/download.
2. Run: ollama pull nomic-embed-text
3. Ask me which folder holds my notes. Split each file at its headings, then into chunks of about 1,200 characters with a 200-character overlap.
4. Embed each chunk with nomic-embed-text through http://localhost:11434/api/embed and store the vectors on this computer.
5. Build a BM25 keyword index over the same chunks.
6. Keep a hash of each chunk. On a refresh, re-embed only chunks whose hash changed.
7. For a question, run both searches and merge the two ranked lists (reciprocal rank fusion). Show me the top five chunks with their file names.
8. Keep everything on this computer. Don't send my notes to any outside service.

Our own numbers here come from our files, checked on 27 October. If a step doesn't work for you, reply to my DM.

— Champ Smith, CYSTEMS

Sources: block.xyz/dessa · techcrunch.com (24 Oct 2026) · ollama.com/download/mac · ollama.com/library/nomic-embed-text · ollama.com/library/qwen3 · formulae.brew.sh openai-whisper · our own search, lead sorter and transcription files, checked 27 Oct 2026.