Architecture
CodeBroker is designed as a robust, multi-stage pipeline designed for scale and precision.
Data Pipeline
Repository
Raw source code files on disk.
Parser
Tree-sitter generates Abstract Syntax Trees.
Semantic Extraction
Identifies declarations, references, and imports.
Flow Analysis
Traces data flow and call graphs.
Relationship Resolution
Links definitions to references across files.
Graph Builder
Constructs the directed semantic graph.
Feature Extraction
Calculates PageRank and centrality metrics.
Embeddings
Generates vector embeddings for nodes.
Retrieval
Traverses graph based on query heuristics.
Context Capsules
Packages subgraph into LLM-ready markdown.