How Call Graphs Give Our LLM Superhuman Context

The bottleneck in AI code review isn’t the LLM—it’s the context. Here’s how we solved it.

The Problem

LLMs are incredibly powerful at understanding code, but they have a critical limitation: token limits. When reviewing a pull request, you can’t just feed the entire codebase into the model.

Traditional approaches:

• ❌ Just the diff: Misses important context
• ❌ Entire repo: Exceeds token limits
• ❌ Random sampling: Loses critical dependencies

Our Solution: Call Graphs

We built a call graph system that maps every function relationship in your codebase. When reviewing a PR, we use this graph to extract only the most relevant context.

How It Works

Instead of guessing what’s relevant, we know exactly which functions are called and which call the changed code.

The Algorithm

1. Parse the codebase using TreeSitter
2. Build dependency graph of all functions
3. Identify changed functions in the PR
4. Traverse graph to find related code
5. Rank by relevance and fit within token budget

Real-World Impact

Before Call Graphs

// AI only sees the diff
async function checkout(cartId: string) {
  const cart = await getCart(cartId);
  return processPayment(cart.total); // ⚠️ AI doesn't know what processPayment does
}

Result: Shallow review, missing critical issues

After Call Graphs

// AI sees the full context
async function checkout(cartId: string) {
  const cart = await getCart(cartId);
  return processPayment(cart.total);
}
 
// AI also knows about:
// - processPayment() implementation
// - chargeStripe() that it calls
// - Error handling in the payment flow
// - How failures propagate

Result: Deep, contextual review that catches real bugs

Performance

Our call graph system is fast:

• Initial build: ~2 seconds for 100K LOC
• Incremental updates: ~200ms per PR
• Graph traversal: ~50ms
• Total overhead: < 3 seconds

The Results

Since implementing call graphs:

• 📈 60% better issue detection
• 🎯 45% fewer false positives
• ⚡ 30% reduction in token usage
• 💪 Superhuman context understanding

Want to experience superhuman code review? Try Mesrai free