nwhitehouse tries a sturdier citation pipeline, then walks it back
An experiment to make AI citations more reliable on a small in-house model hit a wall - but the surrounding polish stuck.
nwhitehouse's fork runs on a tiny tuned model, and one persistent annoyance is that when the AI cites a source, the citation markers don't always render properly. The team tried a sturdier mechanism - instead of asking the model to spit out citations as free-form text, give it a dedicated "tool" it has to call for each citation, on the theory that small models follow structured tool calls more reliably than free-form formatting.
It didn't work. The model happily wrote the citation numbers in its answer but skipped the tool call, so nothing rendered. Forty minutes later the change was reverted to the original approach. What survived: a nicer hover preview on citations (filename, page, a serif pull-quote) and a fix for a small but irritating bug where clicking a second citation in an already-open document didn't scroll to the new spot.
Spotted something wrong? Or know the PR text has fresher detail than the writeup above?