Skip to content
N
Booting Engineering OS
000%
All projects
Building#2 · Core engine

Codebase Intelligence

Deep, queryable understanding of any codebase — grounded answers with citations, living docs, and impact analysis, over API + MCP.

The brain that actually understands your code — searchable, explainable, and queryable by humans and agents alike.

View repoCodebase understanding
Status
Real evaluated RAG-over-code engine working (recall@3 = 100%, MRR 0.818 → 0.864 with re-ranker). Apache-2.0.
Difficulty
High
Time to MVP
2–3 months

Role in the platform

The substrate and technical center of gravity. Master RAG-over-code once, power three products (ContextOS #1 and System Design Assistant #6 reuse it).

Overview

A platform that ingests a codebase and builds rich, retrievable understanding of it: semantic + lexical search, grounded Q&A, automatically maintained documentation and architecture maps, and impact analysis.

It is the RAG-over-code engine — the hardest, most defensible piece of the whole family — exposed as an API and an MCP server so any tool or agent can use it.

The problem

'Talk to your codebase' is validated, hot demand, but most tools are shallow — naive chunking, weak cross-file reasoning, no governance.

Onboarding, code understanding, and impact analysis are universal, expensive pains across large, legacy, and polyrepo codebases.

What it is

An ingest → AST-chunk → embed → hybrid-retrieve → graph-expand → re-rank → cite/answer pipeline backed by pgvector + full-text search and a symbol/dependency graph in Postgres.

Connected via a GitHub App; answers come back grounded with file:line citations, plus auto-generated docs and architecture maps that stay in sync with the code.

Key features

Grounded Q&A with citations

Ask 'how does auth work here?' and get answers grounded in real code with file:line citations — or an honest 'I don't know'.

Hybrid retrieval + re-ranking

AST-aware chunking, semantic + keyword search, import-graph expansion, and a re-ranker for high answer fidelity.

Living docs & architecture maps

Automatically maintained documentation and diagrams derived from the code, so they can't drift.

Impact / blast-radius analysis

'What breaks if I change this function?' across files and repos.

API + MCP server

Every capability is exposed so any tool or agent can query the codebase.

Tech stack

TypeScriptPostgreSQL + pgvectorFull-text searchAST chunking + symbol graphEmbeddings + re-rankerGitHub AppMCP server

Why it's different

  • Over-invests in retrieval quality — evaluated in CI (recall@k, MRR), 'I don't know' over hallucination.
  • API/MCP-first and shared/governed, unlike per-user, editor-bound '@codebase' indexes.
  • Deeper than editor indexes, lighter to adopt than heavy enterprise code-intelligence suites.

Roadmap

MVP

Connect repo, index, semantic + keyword search, grounded Q&A with file:line citations, auto repo summary.

V1

Teams/RBAC, living auto-docs, architecture diagrams, multi-repo, impact analysis.

V2

Agents (codebase Q&A, PR review), MCP server, automation, enterprise controls.

V3

On-prem/VPC, SSO, advanced governance, org-wide knowledge graph.