Skip to content

Smarter, language-agnostic prompt scoring - #2

Open
PiWebswiss wants to merge 1 commit into
ThomasPixelsOf:mainfrom
PiWebswiss:main
Open

Smarter, language-agnostic prompt scoring#2
PiWebswiss wants to merge 1 commit into
ThomasPixelsOf:mainfrom
PiWebswiss:main

Conversation

@PiWebswiss

Copy link
Copy Markdown

Make the complexity heuristic sharper and fix real misfires, so it works consistently whatever language the user writes in.

Make the complexity heuristic sharper and fix real misfires, so it works
consistently whatever language the user writes in.

- scoring.js: match keywords at a word START (Unicode-aware) instead of
  raw substring, killing phantom hits like "just"/"juste" inside "adjusted"
  or "vite" inside "invite". Suffixes still match as light stemming
  (compared->compare, renommer->renomme).
- scoring.js: add structural, language-independent signals — error/stack
  vocabulary + "file.ext:line" references (+3), and multi-step lists of 3+
  numbered/bulleted lines (+2). Pasting a stack trace or a 3-step plan now
  reads as complex instead of scoring zero.
- extension.js: computeProjectBias used the same flawed substring match over
  README prose; route it through the shared countKeywordHits helper.
- package.json: English complexity/simple keyword parity with French
  (understand, complex, "how does", root cause, …; add an import, correct, …).
- scoring.test.js: 5 new tests (word-start matching, accents, stemming,
  error signal, list signal).
- ecoprompt-0.9.0.vsix: repackaged so the bundled code matches source.

Verified: 36/36 tests pass; no regression on reference FR/EN prompts;
end-to-end run fires the expected suggestion.
Sign up for free to join this conversation on GitHub. Already have an account? Sign in to comment

Labels

None yet

Projects

None yet

Development

Successfully merging this pull request may close these issues.

1 participant