RAG
RAG retrieves enterprise documents first, then has a large language model compose answers from that material. It is not “training” another ChatGPT on your entire company file dump.
RAG (Retrieval-Augmented Generation) is the usual enterprise knowledge-base pattern: retrieve passages from PDFs, SOPs, FAQs, or a database, then have a large language model answer from those passages and cite the source. When documents change, rebuild the index—you do not train a new model on the whole company.
Import and parse PDF, Office, or database content.
Chunk, generate embeddings, store in PostgreSQL / pgvector or another vector store.
Retrieve relevant chunks, have the model answer, and show sources. Add reranking when needed.
時程：Small-scope knowledge bases commonly take 4–10 weeks, depending on document quality, permissions, and admin needs.
費用：AI / RAG systems typically from NT$100,000–300,000, adjusted by data volume, permissions, and whether existing systems are integrated.
正式價格於需求確認後提供，以上為參考。
Leaf AI Workspace：掃描 NAS 資料夾做企業文件知識庫、附頁碼回答，並用受控工具查專案／帳款，通知前先人工確認。
說明 RAG 如何把公司 PDF、SOP 與 FAQ 變成可提問的知識庫，以及它和直接使用 ChatGPT 的差別。
整理企業 RAG 知識庫的價格區間、影響報價的因素，以及為什麼不適合在看文件前報死價。
比較把檔案貼上聊天、使用企業方案，以及自建 RAG／API 串接的差異與風險。
RAG (Retrieval-Augmented Generation) finds relevant passages in documents you provide, then asks a large language model to answer only from those passages. It fits policies, SOPs, FAQs, and product docs that change. It is not training a new model on the whole company archive.
General ChatGPT answers from the model’s prior knowledge. RAG searches documents you provide, requires the model to answer from retrieved content, and cites sources. It can still misread a passage—so you must be able to open the original.
Company policies, products, and SOPs change—start with RAG. Fine-tuning suits fixed tone or format, not a living knowledge base.
If the job is answering from existing documents with citations, RAG is the usual approach. Summaries, rewrites, or classification that do not need a source passage can use a general LLM API. Add a controlled Agent when the system must query a database or call tools.
A small knowledge base commonly takes 4–10 weeks. AI / RAG systems are referenced from NT$100,000–300,000. A PoC that only checks whether one document set can be queried can start from NT$50,000. Actual scope depends on document quality, permissions, and existing-system integration.
Yes. The Leaf AI Workspace public demo scans a designated folder, incrementally syncs admin, HR, and care documents, and cites filename and page. Production can mount a real NAS without changing retrieval. You still define who may query which files and who retires outdated documents. Try it at https://leafflow-ai.vercel.app/knowledge
Typical enterprise RAG does not fine-tune or train a model. Documents stay in your retrieval index and only relevant passages are sent to an API at question time. Whether a vendor may retain data depends on the API plan and contract.
Limit the searchable document set by login identity. Filter before answering so department A can’t retrieve department B’s internal files.
Tell us the document types, who may query them, and whether answers must cite sources. We reply with whether RAG, an API, or an Agent fits, plus a PoC scope and cost range.
Book a RAG assessment