A conversational transcript may be difficult for people to read when stored for machines. JSON provides a structured representation using objects, arrays, and primitive values, making conversation data easier for downstream applications to consume. section 12.2.1 section 12.3.2
Structured conversational data becomes more useful when its shape is deliberate:
- Represent a running dialogue as ordered assistant and user entries, with optional metadata such as button choices. section 12.2.1
- Extract only the fields an application needs, such as TaxID, MemberID, and ClaimID. section 12.3.2
- Specify exact key names and value types when another application will read the result. section 12.3.2
Extracting claim information
A transcript can be summarized as {"TaxID": 123456789, "MemberID": 87654321, "ClaimID": 111222333}. Adding an example schema to the prompt reduces unwanted fields, prevents duplicated data, and helps preserve the names expected by the receiving application. section 12.3.2
Listing 12.2 Conversational transcript via the built-in “session history” variable
[{"a":"Welcome to the automated assistant, how can I help?"},{"u":
"Claim status","n":true},{"a":"Ok, claims.\nWhat's your Tax ID?"},{"u":
"123456789"},{"a":"Thank you.\nWhat's the member's eight-digit Member
ID?"},{"u":"87654321"},{"a":"Ok, and what's the member's date of
birth?"},{"u":"12/31/2000"},{"a":"Is the member's name John Doe?\n
option: ["Yes","No"]"},{"u":"Yes"},{"a":"Great!\nWhat's the date of
service for the claim?"},{"u":"2024-02-01"},{"a":"I found your claim
#111222333 from Feb 1, 2024 for the member #87654321 and provider
#123456789. It was paid on 5/23/2023 for $201.83."},
{"u":"representative"}]
Inspect the session-history listing to see a JSON array containing alternating assistant and user messages, collected identifiers, confirmations, and a request for a representative. section 12.2.1
