AST based chunking with vector embedding

Most of the time spent resolving a software defect is not spent writing the fix. It is spent finding

the code that needs fixing. In a mature codebase of hundreds of thousands of lines, a bug report

describes a symptom in the language of the customer — "cancellations fail after payment" —

while the defect itself lives in a method whose name, file, and author the reporter has never

seen. Bridging that gap is slow, depends on tribal knowledge, and scales badly as the codebase

and the team grow.

This note describes an approach that closes the gap automatically. Source code is parsed into

syntactically complete units, converted into mathematical representations of its meaning, and

stored in a searchable index. When a defect is reported, a large language model (LLM) reads

the report, searches that index by meaning, retrieves the most likely culprit code, and proposes

a fix for human review.

For the business, the payoff is a measurable reduction in mean time to resolution, less reliance

on a handful of senior engineers who "know where everything is," and faster onboarding. For

engineering, it is a retrieval system precise enough to be trusted as a first responder to incoming

defects, built on well-understood components and kept honest by human review.