Skip to main content

LLM Response Parsing

3 requirements · 16 scenarios

Unified Quote Processing Pipeline

JR-llm-response-parsing-003
The system SHALL provide a process_quotes_pipeline() function that formats and verifies quotes against source documents.
9 test scenarios
  • Quotes formatted as block or inline JR-llm-response-parsing-003.1
  • Quotes verified against primary source JR-llm-response-parsing-003.2
  • Quotes verified against secondary source JR-llm-response-parsing-003.3
  • Verification stored as data attributes JR-llm-response-parsing-003.4
  • Rendering applies colors from data attributes JR-llm-response-parsing-003.5
  • Mismatch triggers auto-correction for documents JR-llm-response-parsing-003.6
  • Mismatch preserved for email body sources JR-llm-response-parsing-003.7
  • Duplicate quotes tracked individually JR-llm-response-parsing-003.8
  • Corrections threaded through pipeline JR-llm-response-parsing-003.9

Citation Paragraph Auto-Correction

JR-llm-response-parsing-011
When a quote is found in a document/attachment but in a different paragraph than cited, the system SHALL auto-correct the citation to reference the actual paragraph where the quote was found.
4 test scenarios
  • Auto-correct for attachment sources JR-llm-response-parsing-011.1
  • No auto-correct for email body sources JR-llm-response-parsing-011.2
  • Correction data stored in HTML attributes JR-llm-response-parsing-011.3
  • Corrections applied during citation rendering JR-llm-response-parsing-011.4

Citation Eligibility Enforcement

JR-llm-response-parsing-012
The system SHALL restrict LLM citations to sources that were actually provided to the LLM in the current turn.
3 test scenarios
  • Ineligible sources excluded from LLM prompt context JR-llm-response-parsing-012.1
  • Ineligible citations removed or flagged JR-llm-response-parsing-012.2
  • Eligibility is tracked once and reused JR-llm-response-parsing-012.3
All Specs