FeatLens: Feature-Guided Dependency Retrieval for Repository-Level Code Generation

FeatLens connects functional requirements with dependencies dispersed across a repository, reducing the need for multi-step online LLM exploration.

• Builds an offline feature–function index and expands entities along call, use, and ownership relations.
• Combines semantic relevance and structural heuristics to construct a personalized PageRank restart distribution and select context under a fixed budget.
• Makes zero LLM calls during online dependency retrieval. Offline indexing and downstream code generation are separate stages; this is not a claim that the entire pipeline is LLM-free.
• At the same Top-15 budget, improved dependency recall over UniXcoder RAG by 11.2 percentage points on DevEval and 17.4 percentage points on EvoCodeBench.
• On DevEval, achieved 55.03% generation Pass@1 with GPT-5-mini.

Status: Preprint. The results above follow the manuscript experiments reported in my CV.

Preprint: arxiv.org/abs/2609.26480.