Function calculateKeywordMatchScore

src/api/utils/vectorStore.ts:752–836

* Generate combined keyword&pattern score for text matching a specific query intent * Used in post-processing to re-rank results beyond vector similarity

(text: string, query: string)

 * Used in post-processing to re-rank results beyond vector similarity
 */
function calculateKeywordMatchScore(text: string, query: string): number {
  // Lower-case for case-insensitive matching
  const lowerText = text.toLowerCase();
  const lowerQuery = query.toLowerCase();

  let score = 0;

  // Penalize license sections, which are rarely relevant
  if (
    /^#+\s+license\b/im.test(text) ||
    text.toLowerCase().includes("mit license")
  ) {
    score -= 0.3;
  }

  // Penalize badge sections which are usually not informative for queries
  if (
    /\]\(https?:\/\/[^)]*badge[^)]*\)/i.test(text) &&
    text.split("\n").length < 8
  ) {
    score -= 0.2;
  }

  // Boost sections that likely contain actual information
  if (
    /^#+\s+(what is|getting started|introduction|usage|examples|installation)/im.test(
      text,
    )
  ) {
    score += 0.3;
  }

  // Extract terms from query (removing stop words)
  const queryTerms = lowerQuery
    .split(/\W+/)
    .filter((term) => term.length > 2 && !commonWords.has(term));

  // Count term occurrences in text
  for (const term of queryTerms) {
    // Use regex to find whole word matches
    const regex = new RegExp(`\\b${term}\\b`, "gi");
    const matches = lowerText.match(regex) || [];

    // Add score based on frequency
    score += matches.length * 0.05;
  }

  // Boost for heading matches (much higher boost than before)
  const headings = text.match(/#{1,6}\s+([^\n]+)/g) || [];
  for (const heading of headings) {
    const lowerHeading = heading.toLowerCase();
    for (const term of queryTerms) {
      if (lowerHeading.includes(term)) {
        score += 0.25; // Higher boost for term in heading
      }
    }
  }