The Shirt Board has been quietly running for over twenty years. Somewhere in those threads is almost every screen printing question that’s ever been asked, and answered, argued about, answered again better. The problem is finding any of it. Search engines barely touch old forums anymore. So we built a system that lets an AI actually read the whole thing, and prove where every answer came from. °FAI: RAG from the forum we’ve all used.

The technical name for this is retrieval-augmented generation, RAG. The plain name is: the AI looks it up before it talks.

First, we made a copy. Politely.

A small Python program visited every board a guest can see and saved every thread as its own text file, title, board, post count, and a link back to the original at the top, then every post underneath with its author and date intact.

It kept a running list of finished threads, so if it stopped mid-crawl it picked up exactly where it left off. And it waited a second and a half between every request. Slower than a person clicking around. A twenty-year-old forum deserves a polite guest.

The final count: 20,820 threads. 187,100 posts. 57 boards. All of it fits in 138MB, smaller than a couple of videos on your phone.

Then we made it searchable.

The copy went into a small database with a full-text index, the same kind of ranking math (it’s called BM25) that search engines have run on for decades. It’s smart enough to know “cure,” “cured,” and “curing” are the same word.

And here’s the part I care about: it scores individual posts, then adds those scores up per thread. So a real discussion, five printers going back and forth, testing, disagreeing, landing somewhere, outranks a single post that happens to repeat your keyword six times. Substance beats noise. That’s the whole point of the ranking.

A search returns the top threads with highlighted snippets, can be narrowed to one board, and can pull up any full thread on demand.

The AI doesn’t get to make things up.

This is the step where most AI writing goes wrong, so we kept a human in the loop. The AI runs several searches on a topic, not one, several, from different angles. It reads the top threads in full, not just the snippets. Then it drafts an article where every claim traces to a source. Direct quotes link to the exact post they came from, verified links, checked against the archive.

If it’s not in the forum, it doesn’t go in the article. The AI isn’t the expert here. The printers who spent twenty years answering each other are. The AI is the reading aid. So check out Throwback Thursday on Thursdays and see how the RAG is performing.

The honest part

There’s no fancy semantic search yet, no embeddings, none of the vector database stuff the AI world loves to talk about. This is keyword search, the old reliable kind. If it starts missing things because someone said “ghosting” where I searched “dye migration,” that’s the planned upgrade.

 AI didn’t write this archive and it can’t replace it. But it can finally read it, compile it and hand it back in a way that we can still re-experience the learning.

The forum did the knowing. The AI just does the finding.

Tools used:

• Claude (runs the searches, reads the threads, drafts the articles — with a human checking the work)
• Python (requests + BeautifulSoup, the polite scraper)
• SQLite with FTS5 (the small database and full-text index that make 187,100 posts searchable)
• The Shirt Board (twenty years of printers answering each other — the actual source)