Exa Architecture Variants
Overview
Three deployment architectures for Exa neural search at different scales. Each uses real Exa SDK methods: search, searchAndContents, findSimilar, getContents, and answer.
Decision Matrix
| Factor | Direct Search | Cached Search | RAG Pipeline |
|---|
| Volume | < 1K/day | 1K-50K/day | Any volume |
| Latency | 500-2000ms | ~50ms (cached) | 3-8s total |
| Use Case | Simple search UI | Content aggregation | AI answers with citations |
| Complexity | Low | Medium | High |
| Cache Required | No | Yes (Redis/LRU) | Yes |
| Exa Methods | searchAndContents | searchAndContents + cache | All methods |
Instructions
Variant 1: Direct Search Integration
Best for: Adding search to an existing app, < 1K queries/day.
import Exa from "exa-js";
import express from "express";
const app = express();
const exa = new Exa(process.env.EXA_API_KEY);
app.get("/api/search", async (req, res) => {
const query = req.query.q as string;
if (!query) return res.status(400).json({ error: "q required" });
try {
const results = await exa.searchAndContents(query, {
type: "auto",
numResults: 5,
text: { maxCharacters: 500 },
highlights: { maxCharacters: 300, query },
});
res.json(results.results.map(r => ({
title: r.title,
url: r.url,
snippet: r.highlights?.join(" ") || r.text?.substring(0, 200),
score: r.score,
})));
} catch (err: any) {
res.status(err.status || 500).json({ error: err.message });
}
});
Variant 2: Cached Search with Category Profiles
Best for: High-traffic search, 1K-50K queries/day, content discovery.
import Exa from "exa-js";
import { LRUCache } from "lru-cache";
const exa = new Exa(process.env.EXA_API_KEY);
const cache = new LRUCache<string, any>({ max: 5000, ttl: 3600 * 1000 });
const PROFILES = {
news: {
type: "auto" as const,
category: "news" as const,
numResults: 10,
text: { maxCharacters: 500 },
},
research: {
type: "neural" as const,
category: "research paper" as const,
numResults: 10,
text: { maxCharacters: 2000 },
highlights: { maxCharacters: 500 },
},
: {
: ,
: ,
: ,
: { : },
},
};
() {
key = ;
cached = cache.(key);
(cached) cached;
results = exa.(query, [profile]);
cache.(key, results);
results;
}
Variant 3: Full RAG Pipeline
Best for: AI-powered answers, research agents, 50K+ queries/day.
import Exa from "exa-js";
import { LRUCache } from "lru-cache";
const exa = new Exa(process.env.EXA_API_KEY);
const contextCache = new LRUCache<string, any>({ max: 10000, ttl: 7200 * 1000 });
class ExaRAGPipeline {
async gatherContext(question: string, maxSources = 5) {
const cacheKey = question.toLowerCase().trim();
const cached = contextCache.get(cacheKey);
if (cached) return cached;
const results = await exa.searchAndContents(question, {
type: "neural",
numResults: maxSources,
text: { maxCharacters: 2000 },
highlights: { maxCharacters: 500, query: question },
});
contextCache.(cacheKey, results);
results;
}
() {
exa.(topResultUrl, {
: numSimilar,
: { : },
: ,
});
}
() {
results.(
+
+
+
).();
}
() {
answer = exa.(question, { : });
{
: answer.,
: answer..( ({
: r.,
: r.,
})),
};
}
() {
context = .(question, );
expanded = { : [] [] };
(context.[]?.) {
expanded = .(context.[].);
}
allResults = [...context., ...expanded.];
llmContext = .(allResults);
{
: llmContext,
: allResults.,
: allResults.( ({ : r., : r., : r. })),
};
}
}
Scaling Notes
| Architecture | 10 QPS Limit Strategy |
|---|
| Direct | Natural limit: ~864K searches/day at full rate |
| Cached | 50% cache hit = ~1.7M effective searches/day |
| RAG Pipeline | 2-3 API calls per question; cache aggressively |
Error Handling
| Issue | Cause | Solution |
|---|
| Slow search in UI | No caching | Add LRU or Redis cache |
| Stale cached results | Long TTL | Reduce TTL for time-sensitive profiles |
| RAG hallucination | Poor source selection | Use highlights, increase numResults |
| High API costs | No query deduplication | Cache layer deduplicates identical queries |
Resources
Next Steps
For reference architecture details, see exa-reference-architecture.
