Function searchDocumentation

src/api/utils/vectorStore.ts:849–924

(
  owner: string,
  repo: string,
  query: string,
  limit: number = 5,
  vectorize: Vectorize,
)

 * @returns Array of relevant document chunks with scores
 */
export async function searchDocumentation(
  owner: string,
  repo: string,
  query: string,
  limit: number = 5,
  vectorize: Vectorize,
): Promise<Array<{ chunk: string; score: number }>> {
  try {
    // Check if Vectorize is available
    if (!vectorize) {
      console.warn("Vectorize binding not available. Returning empty results.");
      return [];
    }

    // Generate namespace for this repository
    const namespace = getRepoNamespace(owner, repo);
    console.log(`Searching in namespace: ${namespace}`);

    const queryEmbedding = await getEmbeddings(query);

    // Query vectors using Cloudflare Vectorize with namespace
    const results = await vectorize.query(queryEmbedding, {
      topK: limit,
      namespace: namespace, // Use namespace instead of filter
      returnValues: false, // We don't need the vector values back
      filter: {
        timestamp: { $gt: Date.now() - VECTOR_TTL }, // Only keep recent vectors
      },
      returnMetadata: true, // We need the metadata for chunks
    });

    console.log(
      `Found ${results?.matches?.length || 0} results in namespace ${namespace}`,
    );

    if (!results || !results.matches || results.matches.length === 0) {
      console.warn(`No results found in namespace ${namespace}`);
      return [];
    }

    // Enhanced ranking: combine vector similarity with keyword matching
    const enhancedResults = results.matches.map((match) => {
      const metadata = match.metadata as Record<string, any>;
      const chunk = metadata?.chunk || "";

      // Calculate keyword match score
      const keywordScore = calculateKeywordMatchScore(chunk, query);

      // Combine scores (vector similarity + keyword matching)
      // Normalize vector similarity from [-1,1] to [0,1] range if using cosine similarity
      const normalizedVectorScore = (match.score + 1) / 2;

      // Combined score gives weight to both vector similarity and keyword matches
      const combinedScore = normalizedVectorScore * 0.6 + keywordScore * 0.4;

      return {
        chunk,
        vectorScore: match.score,