Experiment 02 / Educational model
The latency budget
A response is only as fast as its critical path. Explore which work can happen at the same time.
MODELED RESPONSE TIME1.45s
300 ms saved by overlapping independent work
Speech to text
0–400 ms
Retrieval
0–300 ms
Language model
400–1200 ms
Text to speech
1200–1450 ms
0 ms1750 ms
What this model assumes
Retrieval can begin independently alongside speech processing. Generation waits for both; speech output waits for generation. No streaming, network variance, retries, or tool execution is modeled. These adjustable values are illustrative, not measured Kaarya or provider performance.
Sequential = STT + retrieval + LLM + TTS
Parallel = max(STT, retrieval) + LLM + TTS