Client is a Traditional Chinese Medicine (TCM) university-level research institute, looking to tap on AI to consolidate, structure, and apply knowledge found in TCM legacy documents and texts.
Knowledge on lumbar pain in TCM is scattered across classical medical texts, textbooks, and journals, with unstructured medical records. Client wishes to compile and construct a specialised disease knowledge graph using AI-driven automated extraction combined with expert confidence-based triage review.
Zhimai is an enterprise-level knowledge solution that configures "Objectives and Key Results" for large language models, designed to address four major challenges in LLM deployment: hallucination (generated content appears plausible but is factually incorrect), knowledge silos (scattered enterprise documents with low utilization), weak complex reasoning (inability to handle multi-layered, strongly correlated business logic), and compliance and security barriers (data cannot be uploaded to the cloud, lack of flexible permission systems).
The core solution is the OKR framework: K (Knowledge Graph, building a private structured fact database for enterprises, where answers must be based on graph triples with visual reasoning paths), O (Ontology constraints, enabling no-code visual definition of knowledge skeletons to prevent LLM overreach), and R (RAG retrieval- augmented generation, with dual-engine collaborative retrieval and multi- dimensional quality assessment).
Five core capabilities: ① Full-cycle closed-loop knowledge processing (multi-source heterogeneous data cleaning, dual-track validation with LLM and NLP, confidence- based review and approval); ② Multi-hop reasoning (strictly following graph topology for step-by-step reasoning, restoring business logic like "because A, therefore B, therefore C," rather than stitching together semantically similar fragments); ③ White- box traceability (five-level traceability evidence: original text location, graph nodes, reasoning chains, one-click report export, operation logs, meeting stringent compliance audit requirements); ④ No-code autonomous control (business experts can independently define knowledge skeletons, upload documents, adjust rules, and release versions, with IT only responsible for initial deployment); ⑤ Debuggable refinement engine with incremental updates and RAG quality assessment (new documents automatically trigger parsing and ingestion, reviewed modifications automatically enter the Few-Shot example library for continuous accuracy improvement without model retraining; provides quantifiable metrics such as recall accuracy, answer match rate, and hallucination rate for monitoring).
The delivery system covers the entire process from agile PoC to continuous operation: establishing a golden test set for measurement standards → MVP integration (prompt engineering, pass rate 70%-85%) → capability injection (RAG mounting private documents, Tool-use connecting business systems) → core optimization (optional SFT fine-tuning/small model distillation) → secure grayscale release → continuous evolution (collecting Bad Cases, regular updates).