PatentTRIZbasedAI20260226110030/Patent-GPT

Implements a dual-path TRIZ classification system (LLM via contradiction matrix or pre-trained ML) feeding a hybrid search layer combining BM25 and cross-encoder reranking against KIPRISplus patent data, with conditional LangGraph evasion loops for design modification when similarity scores exceed thresholds. Outputs are validated through Pydantic structured outputs and formatted as KIPO-compliant DOCX files via python-docx. The architecture separates concerns across FastAPI service layer → orchestration → four testable pipeline stages (TRIZ → Prior Art → Reasoning → Draft), enabling deployment flexibility between CPU-bound ML inference and token-efficient LLM classification.