Pins
An old trick with a short name. Search can miss. The must-see page still goes in.
People have been forcing a known-good page into the prompt for as long as retrieval has been a guess: always-include files, editorial must-reads, a job aid pasted above the question. Help desks pin articles. RAG kits have a box that says include this document every time. Same move. Pin is what we call our version.
Vector search is still an educated guess. Sometimes it ranks the right policy fifth, or grabs a neighbour that uses the same words. Reranking cannot invent a file that never came back. For a few clinic topics we already know the sheet that has to be in the room: needlestick, waste, the policy numbers that must not be faked. So we write that sheet once, in markdown, and we do not leave it to luck.
When the question matches that topic, the proxy pastes the sheet into the prompt. Search still runs on the rest of the library. The model sees the pin and whatever chunks came back. The pin is not a ranking boost. It is not a second search. It is a lookup: this intent, this file, every time.
Force the page in
Help desks pin articles. Editors mark a must-read. RAG kits have an include-this-document box. If you already know the page that must be in the room, you put it there. Ranking luck is not the plan.
We call it a pin
A short markdown cheat sheet. A topic match, then paste. Search still runs. The new bit is treating that as a named, first-class piece of the Faculty setup, not a hidden prompt hack.
Not vector search
Not a score. Not reranking. Not fine-tuning. The sheet is written by us. The match is a lookup table, not a cosine. The library search is a separate step.
How it runs here
- A question hits RAG Proxy.
- The proxy checks a small table: does this look like a pinned topic?
- If it matches, a short known-good markdown file goes into the prompt.
- Search still runs on the dental library. The pin does not replace retrieval.
- The model sees the cheat sheet plus the retrieved chunks. On that topic the sheet is not optional.
In production on Faculty hardware. Shown in the AI Kitchen talk, 27 Aug 2026. Not in the three Perl files on the code page. Those files are the first public snapshot. Pins came after.