Scoped RAG for code review, and why the diff isn't enough

Context

Ask an LLM to review a diff and you get plausible but shallow feedback: it can't see that a renamed function is called in eleven other files, or that your team forbids a pattern the diff happens to use. The missing ingredient is context — but naively stuffing the whole repo into the prompt is expensive and noisy.

The core idea: scope, don't dump

PRism assembles context bounded to what a change actually affects, in two layers.

1. Structural scope via dependency graph

A Tree-sitter pass builds a repo-wide symbol and dependency graph. For a given PR, it resolves the impacted set: the changed symbols plus their call sites elsewhere in the repo.

2. Semantic scope via retrieval

Internal style-guide rules live in a Qdrant vector store. Retrieval is scoped to the change, so only the relevant rules ("we don't use raw SQL in handlers") enter the prompt — not the entire handbook.

Making it fast: cache by commit SHA

Rebuilding the AST and symbol map on every webhook is wasteful because most of a repo is unchanged between commits. PRism caches ASTs and symbol maps in Redis keyed by commit SHA, so unchanged files are never re-parsed.

Outcome

The result is review feedback that reads like it came from someone who knows the codebase: specific to the impacted symbols, aware of internal rules, and cheap enough to run on every PR. Quality is kept honest with a precision/recall pipeline over known-bad PRs.