Query large datasets safely from a realtime voice agent

Goal

Design data-heavy function calls for a Realtime voice agent, filtering by relevance before returning context.

Steps

1. Define fields, size, and filtering for large tool responses; never send the full dataset. [source-openai-serve-large-reference-data-in-realtime-voice]
2. Call a server-side function and return only top-k relevant results and a summary. [source-openai-serve-large-reference-data-in-realtime-voice]
3. Measure latency and context size; ask follow-ups when data is insufficient and avoid silent truncation. [source-openai-serve-large-reference-data-in-realtime-voice]

How to verify

For fixed queries, match `searchDraftProspects` summary IDs to source data, call `getProspectDetails` only on demand, and record payload bytes and time to first audio.

Troubleshooting

If context grows, remove redundant nested fields from detail results; when summary search returns none, ask for filters instead of guessing.