kobotoolbox / kpi / 22687010118 / 2

81%

master: 76%

push

github

fix(qual): handle empty result from LLM DEV-1814 (#6785)

### 📣 Summary
Handle an error resulting from a rare edge case with the LLMs.


### 💭 Notes
On very rare occasions, the LLM will return 'None' instead of a string
answer. This previously resulted in a 500. With this PR we catch it and
use the regular error-handling flow.

### 👀 Preview steps
This can be hard to recreate. The most consistent way I've found to
force the LLM to provide an empty response is to locally set the
MAX_TOKENS to 1.

1. ℹ️ have an account and a project with an audio question called
'audio'
2. Add a submission
3. Add a manual transcript
4. Add a select one question, ideally a nonsensical one
5. "Generate with AI"
6. 🔴 [on main] 500 error
7. 🟢 [on PR] Response completes successfully (result will probably be
empty)

7553 of 11648 branches covered (64.84%)

28837 of 35141 relevant lines covered (82.06%)

0.82 hits per line