What enterprise voice demands of your tool stack

What this post added

This post details how Glean's tool stack was adapted for voice interactions, exposing and addressing limitations in retrieval quality, tool contract detail, and execution reliability. Key improvements include a cleaner, stricter tool stack with better filtering, ranking, and canonical execution paths. The system now prioritizes dynamic skill and tool discovery, using a filter-rank-trim funnel. Lexical weighting was found to be effective for disambiguating repetitive tool names. For voice, tool contracts are now more explicit and inline, including canonical names, server identity, and argument schemas. Error handling has been enhanced with structured retry signals. The execution path for action-oriented work is now more deterministic, with explicit authorization state and confirmed downstream results. The post also highlights how voice exposed existing infrastructure issues like caching and pooling, which were then addressed to improve real-time system reliability.