本研究旨在設計與實作一套代理式知識管理系統，應用於商業情報文件之自動化摘要、產業標籤化與問答查詢。面對企業常見之 PDF 產業文章、商業報告與非結構化文字資料，本研究整合 PyMuPDF、ChromaDB、Ollama 與 gemma:2b 等工具，建立從文件解析、產業分類、文件級入庫、向量檢索到回答生成的完整流程。系統採用文件級檢索設計，將單篇 PDF 視為一筆 document，並於 metadata 中記錄產業類別、來源檔名與檢索單位，以強化來源追溯與知識庫管理。

為比較不同檢索流程之效果，本研究將傳統 RAG 與 Agent RAG 作為主要比較對象。傳統 RAG 不使用產業條件，直接於整體知識庫中進行全庫檢索；Agent RAG 則在檢索前加入產業條件與 metadata filter，以限制檢索範圍並降低跨產業文件干擾。為避免概念混淆，本研究進一步區分 target-category mode 與 routing mode，前者使用標準目標產業驗證 metadata filter 的降噪效果，後者則由模型依問題內容自行判斷產業，以評估端到端代理式路由能力。

實驗結果顯示，系統於單文件入庫測試中達到 96.67% 的產業標籤化正確率，並在多數題目中維持良好的回答語意一致性。在多產業混合知識庫實驗中，Agent RAG target-category mode 可將平均 Noise Ratio 由傳統 RAG 的 0.3611 降至 0.0000，來源命中率由 48.33% 提升至 61.67%，平均回答評分亦由 2.88 提升至 3.13。Routing mode 雖可降低部分跨產業干擾，但其效果仍受模型產業路由正確率影響。整體而言，本研究證明以文件級檢索、產業標籤化與 metadata filter 結合之代理式知識管理系統，能提升多產業商業情報知識庫之檢索範圍控制與實務應用可行性。

This study designs and implements an agentic knowledge management system for automated summarization, industry labeling, and question answering of business intelligence documents. To address the difficulty of organizing PDF-based industry articles, business reports, and other unstructured textual information, the system integrates PyMuPDF, ChromaDB, Ollama, and the gemma:2b local language model into a complete workflow covering document parsing, industry classification, document-level ingestion, vector retrieval, and answer generation.

The study compares traditional Retrieval-Augmented Generation (RAG) with Agent RAG. Traditional RAG retrieves documents from the entire knowledge base without industry constraints, while Agent RAG applies an industry condition and metadata filter before retrieval to narrow the search scope and reduce cross-industry interference. To avoid conceptual ambiguity, Agent RAG is further divided into target-category mode and routing mode. The former uses the standard target industry to examine the denoising effect of metadata filtering, whereas the latter allows the model to infer the query industry and evaluate the end-to-end routing capability of the agentic workflow.

Experimental results show that the system achieved 96.67% industry labeling accuracy in the single-document ingestion test. In the multi-industry mixed knowledge base experiment, Agent RAG in target-category mode reduced the average Noise Ratio from 0.3611 to 0.0000, improved the source hit rate from 48.33% to 61.67%, and increased the average answer score from 2.88 to 3.13. Although routing mode also reduced part of the cross-industry interference, its performance remained constrained by industry routing accuracy. Overall, the proposed system demonstrates that combining document-level retrieval, industry labeling, and metadata filtering can improve retrieval scope control and provide a feasible prototype for business intelligence knowledge management.

[1] M. Alavi and D. E. Leidner, "Review: Knowledge management and knowledge management systems: Conceptual foundations and research issues," MIS Quarterly, vol. 25, no. 1, pp. 107-136, 2001, doi: 10.2307/3250961.
[2] I. Nonaka, "A dynamic theory of organizational knowledge creation," Organization Science, vol. 5, no. 1, pp. 14-37, 1994, doi: 10.1287/orsc.5.1.14.
[3] T. H. Davenport and L. Prusak, Working Knowledge: How Organizations Manage What They Know. Boston, MA, USA: Harvard Business School Press, 1998.
[4] H. Chen, R. H. L. Chiang, and V. C. Storey, "Business intelligence and analytics: From big data to big impact," MIS Quarterly, vol. 36, no. 4, pp. 1165-1188, 2012, doi: 10.2307/41703503.
[5] S. Negash, "Business intelligence," Communications of the Association for Information Systems, vol. 13, no. 1, pp. 177-195, 2004, doi: 10.17705/1CAIS.01315.
[6] R. Sharda, D. Delen, and E. Turban, Business Intelligence, Analytics, and Data Science: A Managerial Perspective, 4th ed. Boston, MA, USA: Pearson, 2018.
[7] A. Vaswani et al., "Attention is all you need," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2017, pp. 5998-6008.
[8] J. Devlin, M.-W. Chang, K. Lee, and K. Toutanova, "BERT: Pre-training of deep bidirectional transformers for language understanding," in Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics: Human Lang. Technol. (NAACL-HLT), 2019, pp. 4171-4186, doi: 10.18653/v1/N19-1423.
[9] T. B. Brown et al., "Language models are few-shot learners," in Proc. 34th Int. Conf. Neural Inf. Process. Syst. (NeurIPS), 2020, pp. 1877-1901.
[10] C. Raffel et al., "Exploring the limits of transfer learning with a unified text-to-text transformer," Journal of Machine Learning Research, vol. 21, no. 140, pp. 1-67, 2020.
[11] M. Lewis et al., "BART: Denoising sequence-to-sequence pre-training for natural language generation, translation, and comprehension," in Proc. 58th Annu. Meeting Assoc. Comput. Linguistics (ACL), 2020, pp. 7871-7880, doi: 10.18653/v1/2020.acl-main.703.
[12] J. Zhang et al., "PEGASUS: Pre-training with extracted gap-sentences for abstractive summarization," in Proc. 37th Int. Conf. Mach. Learn. (ICML), 2020, pp. 11328-11339.
[13] Gemma Team et al., "Gemma: Open Models Based on Gemini Research and Technology," arXiv preprint arXiv:2403.08295, 2024. [Online]. Available: https://arxiv.org/abs/2403.08295.
[14] Ollama, "gemma:2b model library," 2024. [Online]. Available: https://ollama.com/library/gemma:2b. [Accessed: May 13, 2026].
[15] S. Yao et al., "ReAct: Synergizing reasoning and acting in language models," in Proc. 11th Int. Conf. Learn. Represent. (ICLR), 2023.
[16] T. Schick et al., "Toolformer: Language models can teach themselves to use tools," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2023.
[17] J. S. Park et al., "Generative agents: Interactive simulacra of human behavior," in Proc. 36th Annu. ACM Symp. User Interface Softw. Technol. (UIST), 2023, pp. 1-22, doi: 10.1145/3586183.3606763.
[18] Q. Wu et al., "AutoGen: Enabling next-gen LLM applications via multi-agent conversation," arXiv preprint arXiv:2308.08155, 2023. [Online]. Available: https://arxiv.org/abs/2308.08155.
[19] L. Wang et al., "A survey on large language model based autonomous agents," Frontiers of Computer Science, vol. 18, no. 6, 2024, doi: 10.1007/s11704-024-40231-1.
[20] L. Weng, "LLM-powered autonomous agents," Lil'Log, Jun. 2023. [Online]. Available: https://lilianweng.github.io/posts/2023-06-23-agent/. [Accessed: May 13, 2026].
[21] LangChain, "Agents," 2026. [Online]. Available: https://docs.langchain.com/oss/python/langchain/agents. [Accessed: May 13, 2026].
[22] P. Lewis et al., "Retrieval-augmented generation for knowledge-intensive NLP tasks," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2020, pp. 9459-9474.
[23] K. Guu, K. Lee, Z. Tung, P. Pasupat, and M. Chang, "REALM: Retrieval-augmented language model pre-training," in Proc. 37th Int. Conf. Mach. Learn. (ICML), 2020, pp. 3929-3938.
[24] V. Karpukhin et al., "Dense passage retrieval for open-domain question answering," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), 2020, pp. 6769-6781, doi: 10.18653/v1/2020.emnlp-main.550.
[25] G. Izacard and E. Grave, "Leveraging passage retrieval with generative models for open domain question answering," in Proc. 16th Conf. Eur. Chapter Assoc. Comput. Linguistics (EACL), 2021, pp. 874-880, doi: 10.18653/v1/2021.eacl-main.74.
[26] Y. Gao et al., "Retrieval-augmented generation for large language models: A survey," arXiv preprint arXiv:2312.10997, 2023. [Online]. Available: https://arxiv.org/abs/2312.10997.
[27] A. Asai et al., "Self-RAG: Learning to retrieve, generate, and critique through self-reflection," in Proc. 12th Int. Conf. Learn. Represent. (ICLR), 2024.
[28] N. F. Liu et al., "Lost in the middle: How language models use long contexts," Trans. Assoc. Comput. Linguistics, vol. 12, pp. 157-173, 2024, doi: 10.1162/tacl_a_00638.
[29] G. Salton, A. Wong, and C. S. Yang, "A vector space model for automatic indexing," Communications of the ACM, vol. 18, no. 11, pp. 613-620, 1975.
[30] S. E. Robertson and S. Walker, "Some simple effective approximations to the 2-Poisson model for probabilistic weighted retrieval," in Proc. 17th Annu. Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval, 1994, pp. 232-241.
[31] C. D. Manning, P. Raghavan, and H. Schutze, Introduction to Information Retrieval. Cambridge, U.K.: Cambridge University Press, 2008.
[32] N. Reimers and I. Gurevych, "Sentence-BERT: Sentence embeddings using Siamese BERT-networks," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), 2019, pp. 3982-3992, doi: 10.18653/v1/D19-1410.
[33] J. Johnson, M. Douze, and H. Jegou, "Billion-scale similarity search with GPUs," IEEE Transactions on Big Data, vol. 7, no. 3, pp. 535-547, 2021, doi: 10.1109/TBDATA.2019.2921572.
[34] Y. A. Malkov and D. A. Yashunin, "Efficient and robust approximate nearest neighbor search using hierarchical navigable small world graphs," IEEE Transactions on Pattern Analysis and Machine Intelligence, vol. 42, no. 4, pp. 824-836, 2020, doi: 10.1109/TPAMI.2018.2889473.
[35] P. Indyk and R. Motwani, "Approximate nearest neighbors: Towards removing the curse of dimensionality," in Proc. 30th Annu. ACM Symp. Theory Comput. (STOC), 1998, pp. 604-613.
[36] Chroma Team, "Chroma documentation: Metadata filtering," 2026. [Online]. Available: https://docs.trychroma.com/docs/querying-collections/metadata-filtering. [Accessed: May 13, 2026].
[37] deepset, "Haystack documentation: Metadata filtering," 2026. [Online]. Available: https://docs.haystack.deepset.ai/docs/metadata-filtering. [Accessed: May 13, 2026].
[38] LangChain, "Build a RAG agent with LangChain," 2026. [Online]. Available: https://docs.langchain.com/oss/python/langchain/rag. [Accessed: May 13, 2026].
[39] deepset, "Advanced retrieval: Extract metadata from queries to improve retrieval," Haystack Blog, May 2024. [Online]. Available: https://haystack.deepset.ai/blog/extracting-metadata-filter. [Accessed: May 13, 2026].
[40] W. Yin, J. Hay, and D. Roth, "Benchmarking zero-shot text classification: Datasets, evaluation and entailment approach," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), 2019, pp. 3914-3923, doi: 10.18653/v1/D19-1404.
[41] X. Zhang, J. Zhao, and Y. LeCun, "Character-level convolutional networks for text classification," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2015, pp. 649-657.
[42] K. Kowsari et al., "Text classification algorithms: A survey," Information, vol. 10, no. 4, Art. no. 150, 2019, doi: 10.3390/info10040150.
[43] C. N. Silla and A. A. Freitas, "A survey of hierarchical classification across different application domains," Data Mining and Knowledge Discovery, vol. 22, no. 1-2, pp. 31-72, 2011, doi: 10.1007/s10618-010-0175-9.
[44] R. Nallapati, B. Zhou, C. dos Santos, C. Gulcehre, and B. Xiang, "Abstractive text summarization using sequence-to-sequence RNNs and beyond," in Proc. 20th SIGNLL Conf. Comput. Natural Lang. Learning (CoNLL), 2016, pp. 280-290.
[45] A. See, P. J. Liu, and C. D. Manning, "Get to the point: Summarization with pointer-generator networks," in Proc. 55th Annu. Meeting Assoc. Comput. Linguistics (ACL), 2017, pp. 1073-1083, doi: 10.18653/v1/P17-1099.
[46] Y. Liu and M. Lapata, "Text summarization with pretrained encoders," in Proc. Conf. Empirical Methods Natural Lang. Process. (EMNLP), 2019, pp. 3730-3740, doi: 10.18653/v1/D19-1387.
[47] A. Nenkova and K. McKeown, "A survey of text summarization techniques," in Mining Text Data, C. C. Aggarwal and C. Zhai, Eds. Boston, MA, USA: Springer, 2012, pp. 43-76.
[48] C.-Y. Lin, "ROUGE: A package for automatic evaluation of summaries," in Text Summarization Branches Out, 2004, pp. 74-81.
[49] T. Zhang, V. Kishore, F. Wu, K. Q. Weinberger, and Y. Artzi, "BERTScore: Evaluating text generation with BERT," in Proc. 8th Int. Conf. Learn. Represent. (ICLR), 2020.
[50] S. Es, J. James, L. Espinosa-Anke, and S. Schockaert, "RAGAS: Automated evaluation of retrieval augmented generation," in Proc. 18th Conf. Eur. Chapter Assoc. Comput. Linguistics: System Demonstrations (EACL), 2024, pp. 150-158, doi: 10.18653/v1/2024.eacl-demo.16.
[51] L. Zheng et al., "Judging LLM-as-a-judge with MT-Bench and Chatbot Arena," in Proc. Adv. Neural Inf. Process. Syst. (NeurIPS), 2023.
[52] J. Thorne et al., "FEVER: A large-scale dataset for fact extraction and verification," in Proc. Conf. North Amer. Chapter Assoc. Comput. Linguistics: Human Lang. Technol. (NAACL-HLT), 2018, pp. 809-819, doi: 10.18653/v1/N18-1074.
[53] Artifex Software, Inc., "PyMuPDF documentation," 2026. [Online]. Available: https://pymupdf.readthedocs.io. [Accessed: May 13, 2026].
[54] Artifex Software, Inc., "Appendix 1: Details on text extraction," PyMuPDF Documentation, 2026. [Online]. Available: https://pymupdf.readthedocs.io/en/latest/app1.html. [Accessed: May 13, 2026].
[55] Adobe Systems Incorporated, Document Management - Portable Document Format - Part 1: PDF 1.7, ISO 32000-1:2008, 2008.
[56] International Organization for Standardization, Document Management - Portable Document Format - Part 2: PDF 2.0, ISO 32000-2:2020, 2020.
[57] I. Beltagy, M. E. Peters, and A. Cohan, "Longformer: The long-document transformer," arXiv preprint arXiv:2004.05150, 2020. [Online]. Available: https://arxiv.org/abs/2004.05150.
[58] O. Khattab and M. Zaharia, "ColBERT: Efficient and effective passage search via contextualized late interaction over BERT," in Proc. 43rd Int. ACM SIGIR Conf. Res. Develop. Inf. Retrieval, 2020, pp. 39-48.
[59] B. Xu et al., "RECOMP: Improving retrieval-augmented LMs with context compression and selective augmentation," arXiv preprint arXiv:2310.04408, 2023. [Online]. Available: https://arxiv.org/abs/2310.04408.
[60] J. Manyika et al., "The economic potential of generative AI: The next productivity frontier," McKinsey Global Institute, 2023. [Online]. Available: https://www.mckinsey.com/capabilities/mckinsey-digital/our-insights/the-economic-potential-of-generative-ai-the-next-productivity-frontier. [Accessed: May 13, 2026].