Experimental, early beta. Data model and interfaces are unstable; expect breaking changes between releases.

Open source · Self-contained · MIT License

Precise, grounded answers
about your codebase.

Compile your repository into a Datomic knowledge graph. AI agents query structured facts instead of scanning raw source. 1.3× more accurate across 9 repos in 7 languages, with a TF-IDF tier that delivers most of the lift at 3.7× better quality per input token.

Terminal
$ noum ask ./my-repo "Which files are the biggest risk hotspots?"

# iteration 1: querying files-by-complexity (47 results)
# iteration 2: querying commit frequency and authorship
# iteration 3: cross-referencing co-change patterns

Top 5 risk hotspots:
  src/api/middleware.ts    38 commits, 4 authors, very-complex
  src/core/parser.ts       47 commits, 2 authors, very-complex
  src/db/migrations.ts     31 commits, 6 authors, complex
  src/auth/session.ts      28 commits, 3 authors, complex
  src/cli/commands.ts      24 commits, 5 authors, complex

Five Stages, Then Queries

Each stage builds on the last. Read the full pipeline →

ImportGit history & filesdeterministic
EnrichImport graphdeterministic
AnalyzePer-file semanticsmicro / LLM
SynthesizeComponents & archmacro / LLM
EmbedVector search indexdeterministic

Measured on Real Codebases

40 questions per repo, 9 repos, 7 languages. See the full table →

1.3×

More Accurate

Without 47.7% → With 62.0% mean across 9 repos

3.7×

Better Per Token

TF-IDF tier vs full KG

9 repos

Seven Languages

Same 40 questions per repo


30 Seconds to Start

1

Install

curl -sSL https://noumenon.leifericf.com/install | bash

Also via Homebrew, Scoop, Docker.

2

Run the Demo

noum demo

Pre-built knowledge graph. No LLM credentials needed.

3

Ask

noum ask noumenon "Which files are the biggest risk hotspots?"