Function getEmbeddings

src/api/utils/vectorStore.ts:78–131

(text: string)

 * @returns Vector embedding (simplified)
 */
export async function getEmbeddings(text: string): Promise<number[]> {
  // This is an improved embedding function that creates a better vector representation
  // Still simple but designed to create more topical differentiation

  const view = new Float32Array(1024);

  // Extract key terms and topics from the text
  const keywordExtraction = extractKeywords(text);

  // Use a more sophisticated hash function that weights important terms
  const termWeights = keywordExtraction.reduce(
    (acc, item) => {
      acc[item.term] = item.score;
      return acc;
    },
    {} as { [key: string]: number },
  );

  // Fill the vector with values based on term importance and positions
  const terms = Object.keys(termWeights);

  // Fill base vector with simple hash
  for (let i = 0; i < view.length; i++) {
    // Simple hash function for demo purposes
    let hash = 0;
    for (let j = 0; j < text.length; j += 10) {
      // Sample text at intervals
      hash = (hash << 5) - hash + text.charCodeAt(j) + i;
      hash = hash & hash; // Convert to 32bit integer
    }
    // Normalize between -0.5 and 0.5 (base values)
    view[i] = (hash % 100) / 200;
  }

  // Enhance with keyword features
  for (const term of terms) {
    // Use term to seed a portion of the vector
    const weight = termWeights[term];
    const termHash = simpleHash(term);
    const startPos = termHash % 900; // Avoid last section

    // Enhance specific positions based on term
    for (let i = 0; i < Math.min(term.length * 4, 50); i++) {
      const pos = (startPos + i * 3) % 900;
      // Add weighted value based on term importance
      view[pos] += weight * 0.5 * (Math.sin(termHash + i) * 0.5 + 0.5);
    }
  }

  // Normalize vector to unit length (important for cosine similarity)
  normalizeVector(view);

  return Array.from(view);
}

/**
 * Extract keywords and their importance from text