Class QuickFixSuggestionExtractor

java.lang.Object
org.ek9lang.compiler.common.QuickFixSuggestionExtractor
All Implemented Interfaces:
Function<ErrorListener.ErrorDetails, List<Suggestion>>
public final class QuickFixSuggestionExtractor extends Object implements Function<ErrorListener.ErrorDetails, List<Suggestion>>
Single source of truth for "Did you mean?" / quick-fix suggestions across the EK9 compiler, IDE, and AI tooling. Returns a filtered, ranked Suggestion list combining two strategies:
  1. Common-literal / keyword pool — Levenshtein-near matches against a small fixed pool of EK9 keywords, literals, and common type names. Catches user typos (e.g. flase → false, treu → true) that the symbol table can't help with because keywords aren't symbols.
  2. Fuzzy symbol matches — symbol-table search results, FILTERED by a Levenshtein-cost threshold AND a length-diff threshold relative to the offending token. Without these filters the search returns whatever the closest k symbols were even when irrelevant (e.g. flase → Class1/Class2).

Parse-message expected tokens (e.g. ANTLR's "expecting 'defines', '@'") are deliberately NOT included. They're already rendered in the inline typeOfError text the parser produces, so adding a separate "did you mean: '@'?" line is redundant noise. Kept out of all output channels — CLI classic, CLI visual, IDE Problems tooltip, IDE Quick Fix popup, MCP ek9_query_diagnostics.

Results are ranked by:

  1. Character-bag overlap descending — the strongest typo signal (chars in common). For flase, false shares all 5 unique chars (perfect anagram) and ranks first.
  2. Same-length first (length-diff ascending) — typos rarely change length
  3. Cost ascending — Levenshtein closeness as fine-grained tiebreak
  4. Source priority: parse-token > keyword > symbol
  5. Alphabetical (deterministic)

Bag-overlap-primary is correct over cost-primary because the existing EK9 Levenshtein cost metric has a boundary anomaly (insertion cost 1 along the edge but 2 in the body) that under-charges length-mismatch candidates. For flase, raw cost gives case=3 (gaming the boundary) but false=4 — wrong winner. Bag-overlap captures the "shares the same letters" signal that actually matters for typo correction.

Hold an instance; QuickFixSuggestionText is the inline "Did you mean" text built from it.

  • Constructor Details

    • QuickFixSuggestionExtractor

      public QuickFixSuggestionExtractor()

  • Method Details

    • apply

      public List<Suggestion> apply(ErrorListener.ErrorDetails error)
      Build the ranked, filtered suggestion list for an error. Capped at 5 entries.

      Keyword-pool and fuzzy-symbol matches only run when the compiler signalled "this error might be a typo" by attaching a non-null fuzzySearchResults object. Errors like E07370 Unreachable Statement carry a valid likelyOffendingSymbol (the unreachable label) but no fuzzy context — they shouldn't get "Did you mean?" noise. Parse-message tokens always run since they come from the parser's own grammar knowledge.

      Specified by:
      apply in interface Function<ErrorListener.ErrorDetails, List<Suggestion>>