What enterprise voice demands of your tool stack

This post details the engineering challenges and solutions for integrating voice capabilities into Glean's tool stack, highlighting the increased demands on tool selection, execution, and error handling compared to text-based interactions. Key contributions include: emphasizing retrieval quality and precise tool contracts (canonical names, server identity, argument names, schema detail) for voice models with tighter context windows and latency budgets; introducing a filter-rank-trim funnel for dynamic skill and tool discovery; leveraging lexical weighting for disambiguating repetitive tool verbs; implementing structured retry signals for voice validation errors; establishing a canonical execution path for action-oriented work (find skill, resolve tool, validate, execute, confirm); and exposing underlying infrastructure weaknesses (retrieval, initialization, execution-path issues) through voice interactions, leading to improvements in caching, pooling, and offloading setup from the request path.