Draft Launches Local Knowledge Graph Engine for Deterministic Code Understanding
Key Takeaways
- ▸Draft ships a local, deterministic knowledge graph engine that creates JSONL snapshots of code structure without relying on cloud services or embeddings
- ▸The approach replaces probabilistic semantic search with precise structural queries by name, file path, and line number
- ▸Version-controlled graph files enable architecture reviews as part of code reviews, with diffs showing structural changes
Summary
Draft has introduced a knowledge graph engine that takes a fundamentally different approach to helping AI understand code. Rather than sending code to third-party vector databases or embedding services, Draft ships a Node.js binary that creates a local, deterministic knowledge graph stored as plain JSONL files in your repository's draft/graph/ directory. The engine analyzes codebase structure using tree-sitter WASM and generates indexes for Go, Python, TypeScript/JavaScript, C/C++, and protocol buffers, capturing precise information about functions, call edges, imports, and dependencies.
The approach offers several advantages over conventional AI-powered code tooling. By using deterministic structural analysis instead of probabilistic semantic search, the system answers precise queries like "which functions call buildGoIndex?" with exact file paths and line numbers. The plain text JSONL format means the entire knowledge graph can be version-controlled in git, allowing teams to review architectural changes in pull requests and track how code structure evolves over time. Before making changes, developers can query impact analysis to learn exactly how many tests, docs, and configs are affected—eliminating guesswork.
Draft's approach is particularly relevant for regulated industries and on-premises environments where sending proprietary code to cloud services isn't feasible. Finance, healthcare, and defense sectors where data residency and privacy are non-negotiable can now use AI-assisted code understanding without compromising compliance. The engine includes built-in features for identifying architectural hotspots (ranked by complexity) and detecting circular dependencies, delivering insights that traditional code review processes require manual analysis to uncover.
- Blast-radius impact analysis and hotspot detection provide concrete metrics before code changes
- Zero external services and plain-text storage make it viable for regulated industries where data residency is critical
Editorial Opinion
Draft's deterministic, local-first approach is a sharp departure from the cloud-heavy trend in AI-for-code tools, and it's likely to resonate strongly with enterprise and regulated environments. By making code structure a first-class, version-controlled artifact rather than a probabilistic search result, Draft transforms how AI can reason about complex systems. The ability to detect blast radius and hotspots through graph queries rather than manual inspection is genuinely useful, positioning Draft well to capture market segments where privacy, reproducibility, and determinism matter as much as raw AI capability.
