Hi Product Hunt - Roy here, I make Canonn. Canonn started from a failure I could not prompt my way out of. You update a policy, paste the new version into the system prompt, ask the obvious question, and the model answers with the old rule it memorised in training. The correct text is right there in the context. It still loses to the prior. Every "ONLY use the provided context!!!" rulebook I wrote worked in a clean test and fell apart in production, where that rule is one of twelve in a 552-word support prompt. So we stopped writing rulebooks and trained faithfulness into the weights. Canonn R1 is a 9B model with four disciplines: Ground (answer from the data), Extract (pull the exact value), Abstain (say when it is not there), Cite (point at the passage).

Canonn R1

The LLM that answers only from your documents

The LLM that answers only from your documents