CodeCompress AI

The problem

Large codebases are difficult to navigate using standard GREP or RAG-based vector searches that often miss architectural context. Developers waste thousands of tokens feeding entire files into LLMs just to understand basic dependency flows and call hierarchies.

What it does

• Extracts a deterministic knowledge graph using tree-sitter AST parsing across 37 languages with zero token cost for code.
• Maps relationships including calls, imports, and inheritance with explicit confidence tags (EXTRACTED, INFERRED, AMBIGUOUS).
• Performs graph traversal operations like BFS/DFS queries, shortest path calculation, and community detection via the Leiden algorithm.
• Generates interactive HTML visualizations, God-Node reports, and structured exports for Obsidian, Neo4j, or Mermaid.
• Reduces token consumption by up to 71x for large corpora by querying a compact graph instead of raw file context.

Frameworks & tools

Python 3.10+, tree-sitter, NetworkX, Faster-Whisper, and support for Neo4j/FalkorDB exports.

Why this beats prompting it yourself

Standard prompts rely on the LLM's limited context window and fuzzy memory of file structures. This tool provides a persistent, commit-able graph that ensures structural facts are grounded in AST data, providing 100% accuracy for code relationships that RAG might hallucinate.

Use cases

• Onboard onto a massive, undocumented repository by identifying central God-Nodes.
• Map the impact of a breaking change by tracing all EXTRACTED downstream dependencies.
• Lower LLM operational costs by providing the agent with a compressed graph JSON instead of full source files.
• Audit architectural integrity by detecting unexpected circular imports or hidden dependencies.

Known limitations

Token savings are negligible for small repositories under 10 files. Non-code assets like PDFs or images require LLM sub-agents which incur standard token costs.