Reference solution for taking one customer voice-support session through speech capture, request interpretation, supporting knowledge retrieval, optional operational context lookup, grounded response generation, and spoken delivery.
Unit event
Customer Voice Session
CUSTOMER_VOICE_SESSIONOne customer voice-support session taken through speech capture, request interpretation, supporting knowledge retrieval, optional operational context lookup, grounded response generation, and spoken delivery.
Design variant
One realtime audio model both recognizes the customer's speech and speaks the answer, with the same text support pipeline between them. Not a single full-duplex speech-to-speech session: §2 defers that until distinct input and output durations are proven.
Every value this design resolves, and who owns it. Change an editable value and the figures recalculate. The pricing engine answers, not this page: nothing here multiplies anything, and until it answers the figures beside you are the ones you were already looking at, labelled as such.
58 fields
How long the customer spoke. The only quantity the input-audio meter reads, and never the same number as the assistant's speaking time.
How large the transcript is expected to be. A Scenario assumption carried forward as this Activity's output, not a conversion from duration.
Text this recognition step sends as prompt input. Zero in this solution, stated rather than omitted: an offering whose schedule also holds an unconditional text INPUT rate refuses a request that says nothing about tokens. Wave 3 review §2.
Input tokens eligible for cached-input treatment.
Estimated prior-conversation tokens.
Estimated tokens in the current customer message.
Estimated generated planning tokens.
Estimated system/policy prompt tokens.
Estimated token length of the retrieval query.
Embedding input tokens eligible for cached treatment.
Tokens submitted to the embedding model.
Number of candidate documents/chunks returned.
Estimated total tokens across candidate documents/chunks presented to reranking.
Number of candidate documents/chunks scored for the query.
Estimated total candidate-document tokens supplied to reranking.
Number of rerank queries.
Output tokens if the selected rerank commercial path meters them; zero in the pilot default.
Estimated tokens in the rerank query.
Estimated token size of assembled grounded context.
Estimated token length of normalized operational context when present.
Input tokens eligible for cached-input treatment.
Prior conversation tokens.
Current customer-message tokens.
Expected generated response tokens.
Normalized operational context tokens.
Retrieved evidence tokens.
System/policy instruction tokens.
How long the assistant speaks. The only quantity the output-audio meter reads.
The answer this step is given to speak.
Voice and delivery instructions sent with it.
Where the direct unit cost comes from. Expand an operation for the activities inside it and how each one resolved.
Capture the customer's speech and turn it into a request the support pipeline can read.
Transcribe / Recognize Customer Speech
PricedTurn the customer's spoken request into text the support pipeline can interpret.
A pricing request was made and returned an amount.
Interpret the request and plan what the answer needs.
Interpret and Plan Request
PricedInterpret the customer request and produce a compact plan for downstream processing.
A pricing request was made and returned an amount.
Find and assemble the enterprise knowledge the answer is grounded in.
Prepare Retrieval Query
No pricing requestPrepare a concise enterprise-knowledge retrieval query.
This activity does not price. No request was made, so there is no amount — not an amount of zero.
Embed Query
PricedGenerate a vector embedding for the retrieval query.
A pricing request was made and returned an amount.
Search Enterprise Knowledge
No pricing requestSearch enterprise knowledge using the query representation.
This activity does not price. No request was made, so there is no amount — not an amount of zero.
Rerank Candidates
PricedRerank retrieved candidates for relevance to the support query.
A pricing request was made and returned an amount.
Assemble Grounded Context
No pricing requestAssemble selected evidence into a bounded context for final response generation.
This activity does not price. No request was made, so there is no amount — not an amount of zero.
Look up whatever operational record the request needs and normalize it.
Execute Operational Tool
No pricing requestExecute an application/tool lookup required for transactional context.
This activity does not price. No request was made, so there is no amount — not an amount of zero.
Normalize Operational Result
No pricing requestNormalize operational/tool output into compact context for generation.
This activity does not price. No request was made, so there is no amount — not an amount of zero.
Write the grounded answer as text.
Generate Grounded Response
PricedGenerate the final customer-facing response using customer, conversation, operational and retrieved context.
A pricing request was made and returned an amount.
Speak the answer back to the customer.
Synthesize Spoken Response
PricedSpeak the grounded text answer back to the customer.
A pricing request was made and returned an amount.
Direct unit total
0.140898 USDBars show each operation’s share of the stated direct unit total. An operation with no priced activity shows no amount, which is not an amount of zero.
Activity states not exercised by this configuration
The read model reports these so their absence reads as absence rather than as coverage.
Issues
The runtime reported none for this configuration.
Technical detail and reproducibility