instead of indexing the workspace, use gpu accelerated ultra fast greps and fine tuned model, kinda like windsurf fast context. We tried to make it as cost and token efficient as possible while maintaining accuracy and speed.