It starts with a tree sitter, and produces a unique Benzi ID for all symbols in arbitrarily large codebases.
With all callflow and dataflow resolved, the agent simply queries the codebase instead of reading though it using traditional RAG approaches or trying to "rank" results using embedding-space approaches.
Benzi has several bonus features such as syntax+semantic verified model code edits, runtime tracer to track actual execution through the Benzi compiler, context aware model generated repros to fast track testing, etc.
On the benchmarks side, Benzi + DeepSeek V4 Flash scores 78% on SWE-bench Verified (benchmark details on the benchmark page). For comparison, DeepSeek reports 73.7% as the baseline scaffolding number for v4flash and self reports their own score to be 78.6%. HOWEVER. Benzi gets to this parity reading 3x less source code than the deepseek harness. (benchmarks page again)
Please try it out, and let me know what you think!