AI coding agents re-read entire files every time they need to understand your code. On a medium project, that is 45,000 tokens per query when the agent only needs 4,900. We built Code Context Engine, an open-source tool that uses a code knowledge graph with CALLS and IMPORTS edges to serve only the relevant code chunks.

The graph is central to retrieval quality. When a developer asks about a payment flow, the system finds the top-ranked chunks via hybrid vector and BM25 search, then walks CALLS/IMPORTS edges to pull in related functions from other files automatically. This graph expansion consistently surfaces code the vector search alone would miss.

The pipeline: tree-sitter AST parsing creates semantic chunks (functions, classes, modules). Each chunk becomes a node. Static analysis extracts edges (function calls, imports). Hybrid retrieval combines vector similarity with BM25 keyword matching via Reciprocal Rank Fusion. A confidence scorer blends vector distance, keyword match, and recency. File diversity filtering prevents one large file from dominating results.

Everything is stored in SQLite: sqlite-vec for vector search, FTS5 for keyword search, and a graph table for CALLS/IMPORTS edges. We chose SQLite over a dedicated graph database for simplicity (three files, zero infrastructure, 2 MB install vs 217 MB for LanceDB alone).

Benchmarked on FastAPI (53 files, 20 real queries): 94% token savings, 0.90 Recall@10, 0.4ms query latency. The graph expansion step alone improves recall by surfacing related files that pure vector search misses.

Attendees will learn how to model code as a graph, how graph edges improve retrieval beyond vector similarity, and practical trade-offs between SQLite graph tables and dedicated graph databases for developer tooling.

MIT-licensed: github.com/elara-labs/code-context-engine