Extractive summarization builds a shorter, structured representation by selecting values or details directly from a transcript rather than inventing new information. section 12.3.2
A practical pattern asks an LLM to extract specified fields into JSON. Explicit keys, value types, instructions, and worked examples reduce errors such as renaming ClaimNumber to ClaimID or returning numbers as strings. section 12.3.2
Extracting handoff details
Given a claims conversation, the output can retain TaxID, MemberID, and ClaimID as JSON, optionally alongside a short reason for requesting a human representative. Verify every extracted value against the transcript before deployment and during operation. section 12.3.2
Listing 12.9 An extractive summary without examples
<|instruction|>
Read the following conversation transcript between a #1
user and the automated assistant. Extract all IDs in #1
JSON format. #1
<|transcript|> #2
<... human-readable transcript removed for brevity ...>
<|JSON|> #3
{
"TaxID": "123456789", #4
"MemberID": "87654321", #4
"DateOfBirth": "12/31/2000", #4
"Name": "John Doe", #4
"ClaimNumber": "111222333", #4
"DateOfService": "February 1, 2024", #4
"ProviderID": "123456789" #4
} #4
Compare a basic extraction prompt with examples that constrain fields and output formatting. section 12.3.2
