confirming dates is messy, any recommendation to i...
# support
v
I have created an agent to handle restaurant reservations. I instructed the agent to repeat reservation details for confirmation but the spelling of date is messy - I am using the multilingual agent and in the confirmation it mixes French and English. Do you have any recommendation to improve this ?
r
Yeah, this happens a lot with reservation agents, especially with names, emails, and dates. I’d avoid asking the agent to “spell everything back” naturally, because the model can mess up letters or pronounce them weirdly. Better setup: 1. Repeat normal details conversationally - “Just to confirm, that’s a table for 4 tomorrow at 7:30 PM under Vanessa, correct?” 2. Only spell sensitive fields - Names - Email addresses - Phone numbers - Special notes 3. Use phonetic confirmation for names if needed - “I have the name as Vanessa — V as in Victor, A as in Alpha, N as in November… is that correct?” 4. Store the raw value separately - Don’t rely on the spoken confirmation as the final value. - Have the agent extract structured fields like: - guest_name - party_size - reservation_date - reservation_time - phone - notes Then confirm from those fields. Prompt-wise, you can add something like:
Copy code
text
When confirming reservation details, do not spell every word unless the caller’s name, email, or phone number may be unclear.
Confirm reservation details in a natural sentence:
“Just to confirm, I have a table for [party_size] on [date] at [time] under [guest_name]. Is that correct?”
If confirming a name or email, spell it slowly using phonetic letters only when necessary.
Do not invent spellings. If unsure, ask the caller to spell it again.
Usually the best flow is: collect → structure → repeat naturally → ask yes/no → only clarify the field that’s wrong.
v
Thank you Rinshin ! I'll try to implement that !
c
Hey Vanessa, could you please share your organization ID, assistant ID, and call ID so I can investigate further?
v
org ID is : 85f93142-5e60-4220-bf76-e9376081a10b
don't know where to find assistant ID ?
c
Is your query for the assistant "**Amirauté multilingue**"
v
yes : 994b862a-d6aa-4b98-9e7e-c1c9d41a94fa
Please check last call
Also, it would be nice if I could create a workflow where the data collected in the structured output are the ones repeated to caller for confirmation, would that be possible ?
😊
the workflow suggested by Rinshin above is interesting, how could I set it up ?
r
Yes, that workflow is possible conceptually, but I’d separate it into two steps: 1. Extract the reservation details into structured fields Example fields:
Copy code
json
{
"guest_name": "Vanessa",
"party_size": 4,
"reservation_date": "2026-05-21",
"reservation_time": "19:30",
"language": "fr",
"phone": "+...",
"special_notes": "..."
}
2. Confirm only from those structured fields So instead of letting the assistant freely “rephrase” the date/name every time, you tell it to use the collected values as the source of truth. For multilingual agents, I’d add a strict language rule too:
Copy code
text
Always confirm the reservation in the same language the caller is using.
Do not mix French and English during confirmation.
Before finalizing the reservation, summarize only the structured reservation fields:
- guest name
- party size
- reservation date
- reservation time
- phone number if needed
- special notes if any
Confirm naturally, not by spelling the full date.
Example in French:
“Pour confirmer, j’ai une réservation pour [party_size] personnes, le [reservation_date] à [reservation_time], au nom de [guest_name]. C’est bien correct ?”
If the caller says something is wrong, ask only for the incorrect field again.
Do not modify confirmed fields unless the caller corrects them.
For the date issue specifically, I would avoid having the assistant “spell” or creatively say the date. Store the date as a structured value, then format it consistently before confirmation. Example: - Internal value: 2026-05-21 - Spoken confirmation in French: jeudi 21 mai à 19h30 The key is: structured output should be the source of truth, and the confirmation sentence should be generated from that structured output, not from the raw conversation memory. If Vapi’s built-in structured output can’t be referenced reliably inside the same conversation, then the safer setup is: Assistant collects details → tool/webhook receives structured fields → webhook returns a clean confirmation sentence → assistant reads that exact sentence to the caller. That way the model is not deciding how to format the date every time. Your backend/workflow formats it once, in the correct language, and the assistant just reads it.
v
great, thanks for your help !