Selective Context Retrieval

Most AI code review systems are doing too much.

They feed the model the whole codebase and hope it finds the signal.

What actually works:

• parse the repo into AST-based chunks
• retrieve only the most relevant chunks for the PR
• keep the context focused on the change

On one benchmark, that meant:

• 99 chunks down to 15
• 31,000+ characters down to about 3,500
• an 88.9% reduction in context size

And the surprising part?

More context did not mean better review quality.

• No context: the model misses repository patterns
• Full context: too much noise, lower quality
• Selective context: better precision, recall, and F1

The lesson is simple:

In AI code review, relevance beats volume.

How are you deciding what context your LLM actually needs?