Focused LLM Input Reduces Output Tokens by 63% in Code Generation

A new development, dubbed vexp, an MCP server, has demonstrated a significant breakthrough in optimizing Large Language Model (LLM) performance for coding tasks. The core innovation involves pre-indexing a codebase into a dependency graph and then serving only the contextually relevant code snippets to AI coding agents. While the expected benefits included reduced input tokens and faster execution, the most surprising finding was a substantial 63% reduction in output tokens.

Benchmarking conducted on the FastAPI open-source repository, comprising approximately 800 Python files, utilized Claude Sonnet 4.6. The results were compelling: without the dependency graph, tasks averaged around 23 tool calls, consumed 40,000 input tokens, and generated 504 output tokens, costing $0.78 per task. With the graph, tool calls plummeted to 2.3, input tokens dropped to 8,000, and output tokens were reduced to 189. This translated to a 58% cost reduction and a 22% speed improvement, with the 63% output token reduction being an unexpected bonus.

The underlying mechanism for this output token efficiency appears to be a general property of LLMs: noisy, irrelevant input leads to verbose, exploratory output, whereas focused, pre-filtered input results in concise, direct answers. When presented with excessive context, the LLM tends to generate "narration" as it attempts to orient itself within the provided information. By contrast, a pre-filtered, graph-ranked context allows the model to bypass this exploratory phase and proceed directly to generating the solution.

The approach leverages `tree-sitter AST parsing` to construct a dependency graph stored in SQLite. A single MCP tool then takes a task description, traverses the graph, and returns ranked context, providing full source for high-centrality pivot nodes and compact skeletons for supporting code. The savings varied by task type, with code understanding tasks benefiting the most (64% reduction in output tokens) and bug fixes seeing the least, though still significant, reduction (30%). This suggests that tasks requiring more "exploration" by the LLM yield greater efficiency gains from focused input. The project offers a free tier at vexp.dev, running locally without cloud dependencies, making it accessible for developers to experiment with this optimization technique. This discovery underscores the critical importance of intelligent context management in maximizing LLM efficiency and cost-effectiveness for complex applications like code generation.