Pipecat — BAML vs. Vanilla Prompting (Voice Agent)

This repo compares the same voice intake flow implemented two ways:

• Vanilla — free-form system prompt
• BAML — structured spec (slots & states)

Both modes write JSONL logs to ./logs, and eval/metrics.ts computes side-by-side stats (turns, latency, slot coverage).

📊 Latest Metrics

pipecat-baml-vs-vanilla@1.0.0 metrics
npx ts-node --transpile-only eval/metrics.ts

🏁 Conclusion

• Handoff success: Tie (100% both)
• Slot coverage: Tie (100% both)
• Turns: BAML wins (avg ~5 vs 6 for vanilla in our runs)