Semantic search

{
  "vectorize": [
    { "binding": "DOCS", "index_name": "documentation" }
  ]
}

async function indexDocument(document, env, previousChunkIds = []) {
  const chunks = splitIntoChunks(document.text, 800);
  const vectors = [];

  for (let i = 0; i < chunks.length; i += 1) {
    const values = await embed(chunks[i], env.EMBEDDING_API_KEY);
    vectors.push({
      id: `${document.id}:${i}`,
      values,
      namespace: document.tenant_id,
      metadata: {
        document_id: document.id,
        title: document.title,
        chunk: chunks[i],
        locale: document.locale,
      },
    });
  }

  for (let start = 0; start < vectors.length; start += 1000) {
    await env.DOCS.upsert(vectors.slice(start, start + 1000));
  }

  const currentIds = new Set(vectors.map((vector) => vector.id));
  const obsoleteIds = previousChunkIds.filter((id) => !currentIds.has(id));
  for (let start = 0; start < obsoleteIds.length; start += 1000) {
    await env.DOCS.deleteByIds(obsoleteIds.slice(start, start + 1000));
  }

  return [...currentIds];
}

export default {
  async fetch(request, env) {
    const principal = await requirePrincipal(request, env);
    const { query, locale = "en" } = await request.json();
    const queryVector = await embed(query, env.EMBEDDING_API_KEY);

    const matches = await env.DOCS.query(queryVector, {
      topK: 8,
      namespace: principal.tenantId,
      returnMetadata: "all",
      filter: { locale },
    });

    return Response.json(matches);
  },
};