Multi-Notebook Hybrid Search Engine

🖥️ Cloudflare Hybrid Search Implementation

Global Multi-Notebook Hybrid Search Query Implementation — the backend retriever described in Deployment Guide.md. This page's search tab is a browser simulator of the same RRF logic (k = 60, alpha-weighted).

// Cloudflare Global Multi-Notebook Retriever Implementation

export interface HybridSearchQuery {
  query: string;
  notebookId?: string; // Optional: Pass specific notebook ID or omit for global search
  topK?: number;
  alpha?: number;
}

export class CloudflareHybridRetriever {
  constructor(private env: Env) {}

  async search(params: HybridSearchQuery) {
    const topK = params.topK ?? 5;
    const alpha = params.alpha ?? 0.7;

    const ftsQuery = params.notebookId && params.notebookId !== 'ALL'
      ? this.env.DB.prepare(
          `SELECT id, content, rank FROM chunks_fts
           WHERE chunks_fts MATCH ? AND notebook_id = ?
           ORDER BY rank LIMIT ?`
        ).bind(params.query, params.notebookId, topK * 2)
      : this.env.DB.prepare(
          `SELECT id, content, rank FROM chunks_fts
           WHERE chunks_fts MATCH ?
           ORDER BY rank LIMIT ?`
        ).bind(params.query, topK * 2);

    const [aiResponse, ftsResults] = await Promise.all([
      this.env.AI.run('@cf/baai/bge-large-en-v1.5', { text: [params.query] }),
      ftsQuery.all()
    ]);

    const queryVector = aiResponse.data[0];

    const filter = params.notebookId && params.notebookId !== 'ALL'
      ? { notebook_id: params.notebookId }
      : undefined;

    const vectorResults = await this.env.VECTOR_INDEX.query(queryVector, {
      topK: topK * 2,
      filter,
      returnMetadata: 'all'
    });

    const rrfScores = new Map();
    const k = 60;

    vectorResults.matches.forEach((match, index) => {
      const score = alpha * (1 / (k + index + 1));
      rrfScores.set(match.id, (rrfScores.get(match.id) || 0) + score);
    });

    const ftsRows = ftsResults.results;
    ftsRows.forEach((row, index) => {
      const score = (1 - alpha) * (1 / (k + index + 1));
      rrfScores.set(row.id, (rrfScores.get(row.id) || 0) + score);
    });

    const sortedChunkIds = Array.from(rrfScores.entries())
      .sort((a, b) => b[1] - a[1])
      .slice(0, topK)
      .map(([id]) => id);

    if (sortedChunkIds.length === 0) return [];

    const placeholders = sortedChunkIds.map(() => '?').join(',');
    const hydrationQuery = `SELECT id, notebook_id, doc_id, content, section_path FROM chunks WHERE id IN (${placeholders})`;

    const { results: finalChunks } = await this.env.DB.prepare(hydrationQuery)
      .bind(...sortedChunkIds)
      .all();

    const chunkMap = new Map(finalChunks.map((c) => [c.id, c]));
    return sortedChunkIds.map(id => chunkMap.get(id)).filter(Boolean);
  }
}