exa-performance-tuning

import Exa from "exa-js";
 
const exa = new Exa(process.env.EXA_API_KEY);
 
function selectSearchType(latencyBudgetMs: number) {
  if (latencyBudgetMs < 200) return "instant";
  if (latencyBudgetMs 

// Each content option adds latency. Only request what you need.
 
// Fastest: metadata only (no content retrieval)
const metadataOnly = await exa.search("query", { numResults: 5 });
 
// Medium: highlights only (much smaller than full text)
const highlightsOnly = await exa.searchAndContents("query", {
  numResults: 5,

import { LRUCache } from "lru-cache";
 
const searchCache = new LRUCache<string, any>({
  max: 5000,
  ttl: 2 * 3600 * 1000, // 2-hour TTL
});
 
async function cachedSearch(query

// Run independent queries concurrently instead of sequentially
async function parallelSearch(queries: string[]) {
  const searches = queries.map(q =>
    cachedSearch(q, { type: "auto", numResults: 3 })
  );
  return Promise.all(searches);
  // 3 parallel searches: ~600ms total (limited by slowest)
  // 3 sequential searches: ~1800ms total

// Phase 1: Fast search for URLs only
// Phase 2: Selective content retrieval for top results only
async function twoPhaseSearch(query: string) {
  // Phase 1: metadata only (fast)
  const results = await exa.search(query, { type: "auto", numResults: 10 });
 
  // Phase 2: get content only for top 3 results
  const topUrls = results.results.

function normalizeQuery(query: string): string {
  return query
    .toLowerCase()
    .trim()
    .replace(/\s+/g, " ")       // collapse whitespace
    .replace(/