code-graph-llm

Builds a compact abstract syntax tree (AST) representation of codebases across multiple programming languages without language-specific parsers. Uses a unified graph schema to represent code structure (functions, classes, imports, dependencies) as nodes and edges, enabling consistent analysis regardless of source language. The graph is serialized into a compact format optimized for LLM token consumption.

Converts the constructed code graph into a compact, LLM-friendly text representation that minimizes token count while preserving semantic relationships. Uses techniques like symbol deduplication, hierarchical summarization, and selective edge inclusion to create a serialized format that fits within LLM context windows. The output is optimized for both readability and token efficiency, enabling larger codebases to fit in a single prompt.

Automatically identifies and maps all import statements, module dependencies, and inter-file references within a codebase, building a directed graph of dependencies. Handles multiple import syntaxes (ES6 imports, CommonJS require, Python imports, Go imports, etc.) through pattern matching and heuristic analysis. Produces a queryable dependency graph that reveals code coupling, circular dependencies, and module boundaries without executing code.

Parses and extracts function/method signatures, class definitions, and their metadata (parameters, return types, visibility modifiers, decorators) from source code across multiple languages. Uses regex-based pattern matching and lightweight AST-like analysis to identify callable entities and their interfaces without full semantic parsing. Stores signatures in a queryable format that enables LLMs to understand the public API surface of code modules.

Creates an in-memory or persistent index of the code graph that enables fast queries for specific symbols, functions, files, or relationships. Supports queries like 'find all callers of function X', 'list all files importing module Y', or 'get the dependency chain from A to B'. Uses hash maps, adjacency lists, or similar data structures for O(1) or O(log n) lookup performance. Enables LLM agents to dynamically retrieve relevant code context based on user queries.

Generates human-readable summaries and documentation from the code graph by combining function signatures, dependency information, and structural metadata. Creates markdown or HTML documentation that describes module purposes, public APIs, and inter-module relationships. Uses the graph structure to automatically organize documentation by module hierarchy and dependency chains, reducing manual documentation effort.

Identifies common code patterns and idioms across multiple programming languages by analyzing the code graph for recurring structural motifs (e.g., factory patterns, dependency injection, middleware chains). Uses heuristic matching on function signatures, class hierarchies, and call patterns to detect design patterns without language-specific semantic analysis. Enables LLMs to understand architectural patterns and suggest refactorings based on pattern recognition.

Tracks changes to the codebase between versions by comparing code graphs and identifying added, modified, or removed functions, classes, imports, and dependencies. Produces a delta representation showing what changed in the code structure without requiring full re-analysis. Enables LLM agents to understand code evolution and generate change summaries or migration guides.

Dynamically selects and prioritizes code context based on LLM token budget and query relevance, using the code graph to identify the most important symbols and relationships for a given task. Implements strategies like relevance ranking, hierarchical summarization, and selective edge inclusion to fit the most informative context within token limits. Adapts context selection based on LLM model capabilities and token limits (e.g., GPT-4 vs Claude vs open-source models).