Exa Cost Tuning

Overview

Reduce Exa AI search API costs by implementing caching, choosing the right search type, and limiting result count per query. Exa charges per search with costs varying by plan tier.

Prerequisites

• Exa API account with usage dashboard access

• Understanding of search patterns in your application

• Cache infrastructure (Redis or in-memory)

Instructions

Step 1: Implement Query-Level Caching

import { LRUCache } from 'lru-cache';

const searchCache = new LRUCache<string, any>({ max: 5000, ttl: 3600_000 }); // 1hr TTL # 5000: 5 seconds in ms

async function cachedSearch(query: string, options: any) { const cacheKey = ${query}:${options.type}:${options.numResults}; const cached = searchCache.get(cacheKey); if (cached) return cached;

const results = await exa.search(query, options); searchCache.set(cacheKey, results); return results; } // Typical cache hit rate for RAG: 40-60%, directly reducing search costs by half

Step 2: Minimize Results Per Query

// Don't fetch more results than you'll actually use const SEARCH_CONFIGS: Record<string, { numResults: number; type: string }> = { 'rag-context': { numResults: 3, type: 'neural' }, // RAG only needs top 3 'research-deep': { numResults: 10, type: 'neural' }, // Research needs more 'fact-check': { numResults: 5, type: 'keyword' }, // Exact match is cheaper 'autocomplete': { numResults: 3, type: 'keyword' }, // Quick suggestions };

Step 3: Use Keyword Search When Appropriate

set -euo pipefail

Neural search: best for semantic/conceptual queries (more expensive)

curl -X POST https://api.exa.ai/search
-H "x-api-key: $EXA_API_KEY"
-d '{"query": "best practices for microservices", "type": "neural", "numResults": 5}'

Keyword search: best for specific terms/names (cheaper, faster)

curl -X POST https://api.exa.ai/search
-H "x-api-key: $EXA_API_KEY"
-d '{"query": "RFC 9110 HTTP semantics", "type": "keyword", "numResults": 3}' # 9110 = configured value

Step 4: Deduplicate Searches in Batch Pipelines

// If processing 1000 documents, many will need similar context searches # 1000: 1 second in ms function deduplicateSearches(queries: string[]): string[] { const seen = new Set<string>(); return queries.filter(q => { const normalized = q.toLowerCase().trim(); if (seen.has(normalized)) return false; seen.add(normalized); return true; }); } // Typical dedup rate: 20-40% for batch processing pipelines

Step 5: Monitor and Budget

set -euo pipefail

Check remaining budget and project monthly cost

curl -s https://api.exa.ai/v1/usage
-H "x-api-key: $EXA_API_KEY" |
jq '{ searches_today: .searches_today, monthly_total: .searches_this_month, monthly_limit: .monthly_limit, utilization_pct: (.searches_this_month / .monthly_limit * 100), days_remaining: (30 - (.searches_this_month / .monthly_limit * 30)) }'

Error Handling

Issue Cause Solution

Monthly limit hit early Uncached pipeline queries Add query caching, expect 40%+ savings

High cost per useful result numResults too high Reduce to 3-5 for most use cases

Cache hit rate low Highly variable queries Normalize queries before caching

Budget spike from batch job No search deduplication Deduplicate queries before batch execution

Examples

Basic usage: Apply exa cost tuning to a standard project setup with default configuration options.

Advanced scenario: Customize exa cost tuning for production environments with multiple constraints and team-specific requirements.

Output

• Configuration files or code changes applied to the project

• Validation report confirming correct implementation

• Summary of changes made and their rationale

Resources

• Official monitoring documentation

• Community best practices and patterns

• Related skills in this plugin pack