Spaces:
Sleeping
Sleeping
Commit History
-
fix: normalizer over-expansion + stopword keyword filter + intent/program coverage
- normalizer: remove single-word program expansions that corrupted college names before the resolver ran ('commerce'→B.COM., 'law'/'law college'→Bachelor of Laws, 'arts'→Bachelor of Arts, 'science'→B.SC., 'computer'→BCA, 'teaching'→B.ED.) These ran BEFORE college resolver, renaming 'Gujarat Commerce College' → 'Gujarat B.COM. College' and breaking fuzzy match - normalizer: add Hindi phrases (ke baare mein→'', milta hai→available, kitne saal→how many years, kahan milta→where available) - fuzzy_matcher: filter all-stopword keywords from keyword map during build to prevent 'gls college' matching 'girls college' keyword (partial_ratio=95) - query_planner: add plain English words to _PROGRAM_ALIASES since normalizer no longer expands them: commerce, arts/art, science, teaching, computer - field_filter: add courses intent keywords: batch, evening, morning, year, years, semester, semesters, saal, kahan, where, milta, milega Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic…6f30221 -
feat: Hindi/Gujarati query fixes + smart result count + deduplication
Normalizer: - Add multi-word Hindi phrase mappings: ke liye→for, ke sath→with, ke under→under, se kam→less than, se jyada→more than, etc. - Strip question-word noise: kya hai, kitni hai, kitna, kaunse, koi - Add ladkiyon/ladkon ke liye phrases for gender-aware queries Search engine: - _dedup_by_college(): collapse multiple rows for same college into one (highest-scoring kept) before returning results - _smart_cutoff(): cap results at min(4, top_k) with score-gap detection — returns 1-2 for very specific queries (score ≥ 0.88), cuts at natural score drop otherwise - Dedup runs before smart cutoff so gap detection sees unique colleges Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>f9ead8e -
initial deploy
6fc17e7
