Most AI chat apps make a quiet assumption that costs them a lot: one model is good enough for everything. It isn't.
When I started building Chymera, I wanted to fix that. The idea was simple — instead of locking the user into a single LLM, the system should figure out what kind of question is being asked and send it to the model best suited to answer it.
This is the story of how I built that routing layer, what I got wrong the first time, and what the working version actually looks like.
The Problem With Single-Model Architectures
Every major AI chat product — ChatGPT, Claude, Gemini — lets you switch models manually. But users don't think in terms of models. They just ask questions. The mental overhead of "hmm, should I use GPT-4o or o1 for this?" is friction that shouldn't exist.
Beyond UX, there's a real capability argument. Llama 3.3 70B via Groq is exceptional at code generation, while Qwen QwQ 32B has unusually strong multi-step reasoning. Gemini 2.5 Flash is fast and has native tool use that pairs well with live web search.
No single model wins every category. So why force users to pick?
Designing the Classifier
The first instinct was to use an LLM to classify intent. Send the query to a lightweight model, get back a category, then route accordingly. I tried this briefly and dropped it immediately. The latency was unacceptable — you're adding a full round trip before the actual answer even starts.
The solution was a rule-based classifier written in plain JavaScript. No API call, no model inference, zero latency. It runs synchronously before any request goes out.
The classifier returns one of seven categories: chitchat, coding, reasoning, creative, search, factual, or general.
How It Works
function classifyQuery(query) {
const q = query.toLowerCase().trim();
const words = q.split(/\s+/);
const len = words.length;
// Chitchat: short messages matching known conversational phrases
const cleanQ = q.replace(/[.,!?]/g, '');
const isChitchat = len <= 5 && chitchatPhrases.some(p =>
new RegExp(`\\b${p}\\b`, 'i').test(cleanQ) || cleanQ === p
);
if (isChitchat) return 'chitchat';
// Search: time-sensitive signals win before anything else
if (any(q, realtimeSignals)) return 'search';
// Recent year + question word = almost always needs live data
const hasRecentYear = /\b(2024|2025|2026)\b/.test(q);
const hasQuestionWord = any(q, ['what', 'who', 'when', 'latest', 'current', 'now']);
if (hasRecentYear && hasQuestionWord && len >= 4) return 'search';
// Reasoning, coding, creative, factual — keyword matching in priority order
if (any(q, reasoningSignals)) return 'reasoning';
if (any(q, codingSignals)) return 'coding';
if (any(q, creativeSignals)) return 'creative';
if (len >= 3 && any(q, factualSignals)) return 'factual';
return 'general';
}
function any(text, keywords) {
return keywords.some(k => text.includes(k));
}
The ordering matters more than the keyword lists. Search detection runs before factual detection — because "what is the current price of ETH" needs live data even though it matches factual patterns (what is). Getting this priority order wrong was my first significant bug.
Mapping Categories to Models
Once the classifier returns a category, model selection is a simple switch:
function getModel(queryType) {
switch (queryType) {
case 'coding': return 'llama-3.3-70b-versatile';
case 'reasoning': return 'qwen-qwq-32b-preview';
case 'creative': return 'gemini-2.5-flash';
case 'search': return 'llama-3.3-70b-versatile';
case 'factual': return 'llama-3.3-70b-versatile';
case 'chitchat': return 'llama-3.3-70b-versatile';
default: return 'gemini-2.5-flash';
}
}
Groq handles Llama and Qwen. Gemini 2.5 Flash runs through Google's Generative AI SDK. The two providers need different streaming implementations, so I extracted a shared aiCore.js module that normalises both into a single SSE pipeline. The chat handler doesn't need to know which provider it's talking to.
The Rate Limit Problem
Running on multiple free-tier API keys creates an obvious problem: 429 errors. The fix was a key rotation pool:
function getKeyPool(prefix) {
const keys = [];
for (let i = 1; i <= 10; i++) {
const k = process.env[`${prefix}_${i}`];
if (k) keys.push(k);
}
if (process.env[prefix]) keys.push(process.env[prefix]);
return [...new Set(keys)].filter(Boolean);
}
let groqIdx = 0;
function nextGroqKey() {
const key = groqKeys[groqIdx % groqKeys.length];
groqIdx++;
return key;
}
Each request picks the next key via round-robin. If a key returns 429, the request retries with the next one. With 10 keys in the pool, the effective rate limit headroom multiplies by pool size.
What Runs Where
The architecture has three independently deployed tiers:
Frontend — React 18 SPA on Netlify. Handles auth state, renders the chat UI, consumes the SSE stream token by token.
API — Express on Railway. Runs the classifier, selects the model, manages the key pool, and pipes the LLM stream back via Server-Sent Events.
Data layer — Supabase for PostgreSQL and auth. Row-level security means each user can only read their own chat history — one policy handles it without any application-level filtering.
Mem0 sits alongside the API and handles semantic memory. Before each non-trivial request, the system searches for relevant memories from past sessions and injects them into the system prompt. The user never needs to re-explain who they are or what they're working on.
What I'd Do Differently
1. Handle ambiguity with confidence scores
Queries like "is Next.js better than Remix?" match reasoning signals but are also partly factual. The current system picks one and commits. A confidence score with fallback logic would handle edge cases better.
2. Improve chitchat detection on longer messages
The len <= 5 check prevents false positives but a longer casual message doesn't classify cleanly. A separate tone detector would help here.
3. Stateful key rotation
The round-robin index resets on every serverless cold start. A distributed counter in Redis would distribute load more accurately across concurrent requests.
Try It
Chymera is live here:
Ask it something technical, then ask it something that needs a live answer — you'll see the model badge in the UI change between responses. The classifier runs in under a millisecond and the routing is completely invisible to the user.
If you've built something similar or approached the routing problem differently, I'd like to hear how.