Exa Cost Tuning

import Exa from "exa-js";

const exa = new Exa(process.env.EXA_API_KEY);

// Define cost tiers per use case
const SEARCH_PROFILES = {
  // Cheapest: metadata-only keyword search
  "autocomplete": { type: "instant" as const, numResults: 3 },

  // Low cost: fast search with minimal content
  "quick-lookup": { type: "fast" as const, numResults: 3 },

  // Medium: balanced search for RAG
  "rag-context": {
    type: "auto" as const,
    numResults: 5,
    text: { maxCharacters: 1000 },
  },

  // Higher cost: deep research
  "deep-research": {
    type: "neural" as const,
    numResults: 10,
    text: { maxCharacters: 3000 },
    highlights: { maxCharacters: 500 },
  },
};

async function costAwareSearch(
  query: string,
  profile: keyof typeof SEARCH_PROFILES
) {
  const config = SEARCH_PROFILES[profile];
  if ("text" in config || "highlights" in config) {
    return exa.searchAndContents(query, config);
  }
  return exa.search(query, config);
}

import { LRUCache } from "lru-cache";

const searchCache = new LRUCache<string, any>({
  max: 5000,
  ttl: 3600 * 1000, // 1-hour TTL
});

async function cachedSearch(query: string, opts: any) {
  const key = `${query.toLowerCase().trim()}:${opts.type}:${opts.numResults}`;
  const cached = searchCache.get(key);
  if (cached) return cached;

  const results = await exa.searchAndContents(query, opts);
  searchCache.set(key, results);
  return results;
}
// Typical RAG cache hit rate: 40-60%, directly cutting costs in half

function deduplicateQueries(queries: string[]): string[] {
  const seen = new Set<string>();
  return queries.filter(q => {
    const normalized = q.toLowerCase().trim().replace(/\s+/g, " ");
    if (seen.has(normalized)) return false;
    seen.add(normalized);
    return true;
  });
}

// Before batch processing, deduplicate
const uniqueQueries = deduplicateQueries(allQueries);
console.log(`Deduped: ${allQueries.length} → ${uniqueQueries.length} queries`);
// Typical dedup rate: 20-40% for batch processing

// Neural search: best for semantic/conceptual queries (more expensive)
// Keyword search: best for specific terms/names (cheaper, faster)

function selectCostEffectiveType(query: string): "neural" | "keyword" | "auto" {
  // Use keyword for exact lookups
  if (query.match(/^https?:\/\//)) return "keyword";     // URL lookup
  if (query.match(/^[A-Z][a-z]+ [A-Z]/)) return "keyword"; // Proper nouns
  if (query.includes('"')) return "keyword";               // Quoted terms

  // Use neural for conceptual queries
  if (query.split(" ").length > 5) return "neural";
  return "auto"; // Let Exa decide for ambiguous queries
}

set -euo pipefail
# Check API key usage
curl -s https://api.exa.ai/v1/usage \
  -H "x-api-key: $EXA_API_KEY" | \
  python3 -c "
import json, sys
d = json.load(sys.stdin)
print(f'Searches today: {d.get(\"searches_today\", \"N/A\")}')
print(f'Monthly total: {d.get(\"searches_this_month\", \"N/A\")}')
print(f'Monthly limit: {d.get(\"monthly_limit\", \"N/A\")}')
" 2>/dev/null || echo "Usage endpoint not available"

// Application-level budget tracking
class ExaBudgetTracker {
  private searchCount = 0;
  private dailyLimit: number;

  constructor(dailyLimit = 1000) {
    this.dailyLimit = dailyLimit;
  }

  async search(exa: Exa, query: string, opts: any) {
    if (this.searchCount >= this.dailyLimit) {
      throw new Error(`Daily Exa budget exceeded (${this.dailyLimit} searches)`);
    }
    this.searchCount++;
    return exa.search(query, opts);
  }

  getUsage() {
    return {
      used: this.searchCount,
      remaining: this.dailyLimit - this.searchCount,
      utilization: `${((this.searchCount / this.dailyLimit) * 100).toFixed(1)}%`,
    };
  }
}