Codebase Copilot
Ask anything about a repository and get an answer in which every claim points to the file and lines it came from.
Problem
Joining an unfamiliar project means finding where the business rules live and how flows cross the layers. Asking an LLM without controlled context produces answers that sound convincing but cannot be checked.
What I built
- Complete, local-first RAG pipeline: safe import, incremental snapshot-based indexing, hybrid search and citation validation.
- Three-channel hybrid search — lexical (with an identifier tokenizer,
getUserById→get user by id), vector (pgvector) and file path — combined with Reciprocal Rank Fusion. - Verifiable citations: every
[E1],[E2]… is checked against the excerpts actually sent to the model, and invalid citations are dropped. Without enough evidence, the tool says it does not know. - Measured quality: custom harnesses for Recall@k, MRR, citation validity, abstention and cost over versioned corpora.
Engineering highlights
- No imported code is ever executed
- API bound to
127.0.0.1with a strict CSP - WCAG 2.2 AA
- ⌘K command palette
- Testcontainers, Vitest and Playwright with visual regression
- OpenAPI contract with generated frontend types
Stack
- Java 25
- Spring Boot 4.1
- PostgreSQL 17 + pgvector
- React 19
- TypeScript
- Vite
- TanStack Query
- Radix UI
- OpenAI SDK
