Problem setup: I’m trying to extract three things from GitHub issues: symptom, mechanism, and failure. Right now, I use an LLM to pull out phrases and then apply deterministic rules to filter and classify them. What’s going wrong: This setup is getting messy the LLM output is inconsistent, the rules are brittle, and fixing one case often breaks another. I also see cases where important signals are missed entirely. Constraints: I’m working with a...
Originally asked by Aman Chadha on Stack Exchange (ai) (CC BY-SA 4.0).
