Data Extraction and Structured Output
Travel Booking Information Extraction Challenge: Extract structured information from conversational data while handling missing information appropriately and avoiding hallucinations. Scenario: Processing travel booking conversations to extract key booking details in a structured format. Performance Comparison: Chat Model We evaluate the performance of a state-of-the-art chat model in two configurations:
- Chat model alone (baseline)
- Chat model enhanced with AI21 Maestro
Baseline Chat Model Output (Without AI21 Maestro)
Despite receiving explicit instructions, the chat model fails to properly handle partial information and makes unauthorized inferences: Input PromptIssues Identified
- Instruction Violation: Added year “2023” despite explicit instruction not to infer
- Compliance Failure: Ignored “fill with NA” requirement for incomplete dates
AI21 Maestro Performance
Input Prompt- If any field is not mentioned or can’t be fully extracted from the conversation, fill it with “NA”
- Do not invent, infer, or assume any details that are not explicitly stated in the conversation.
- return a bulleted list where each field is a bullet