The linked post details my engineering solution: a 'Generalizer-Judge-Surgeon' pipeline to remove hallucinations from legislation.
Using a "dumber/faster" model (Flash Lite) strictly as a Judge (comparing Source vs Summary) worked better than asking initial model to self-verify. It acts like a semantic regex.
Call for Contributors:
The project is Source Available. I have the core architecture ready for scaling, but I need locals to help build the specific data adapters/scrapers for their parliaments. If you want to help fix the "legislative black hole" in your country, check the repo.
Happy to answer questions about the prompt engineering or the architecture.
Using a "dumber/faster" model (Flash Lite) strictly as a Judge (comparing Source vs Summary) worked better than asking initial model to self-verify. It acts like a semantic regex.
Call for Contributors: The project is Source Available. I have the core architecture ready for scaling, but I need locals to help build the specific data adapters/scrapers for their parliaments. If you want to help fix the "legislative black hole" in your country, check the repo.
Happy to answer questions about the prompt engineering or the architecture.