生成式AI輔助檔案銷毀智慧審查

Generative AI-Assisted Records Destruction Smart Review

國家發展委員會檔案管理局過去使用分辨式AI技術進行檔案銷毀審查，但因訓練資料不足，模型準確率受限。生成式AI的興起為此帶來突破，透過生成式AI可透過上下文學習及少量資料提示的特性，提高銷毀審查的準確性並推動流程自動化。此外，生成式AI也讓經驗傳承系統化可以有效落實。
生成式AI導入的挑戰之一是避免「幻覺」問題，需透過微調（Fine-Tuned）和中間層架構（檢索擴增生成RAG）來提升準確率和實用性，包含向量化技術、語意搜尋及提示工程等，有效整合生成式AI和內部系統。並且，將檔案銷毀規範和案例向量化存入資料庫，進行檢索擴增生成後，組成提示語（Prompt），由生成式AI產生審查建議，整體流程將實現從檔案銷毀目錄彙送、AI智慧審查到最終核准銷毀檔案目錄下架自動化，並持續通過回饋提升AI準確性。

The National Archives Administration of National Development Council had used“Discriminative”AI technology for records destruction review, but due to insufficient training data, the accuracy of trained model inferences was limited. The rise of“Generative”AI brings breakthroughs to this. By leveraging In-Context-Learning and Few-Shot-Prompting, Generative AI can improve the accuracy of destruction reviews and enable the automation of the process. Additionally, it enables systematic transfer and effective implementation of expertise.
One of the challenges in implementing Generative AI is to prevent the issue of“hallucinations.”To improve accuracy and practicality, Retrieval Augmented Generation (RAG) architectures must be applied to the process, including vectorization techniques, semantic search, and prompt engineering. It can effectively integrate Generative AI with internal systems. After the regulations and cases of records destruction are vectorized into database, RAG is used to construct prompts, allowing Generative AI to provide reviews.
The overall process will be fully automated─from the compilation of records destruction catalogs and AI-powered intelligent review, to the final approval and removal of record destruction catalogs. The system will continuously improve AI accuracy through feedback mechanisms.