AI
-
RAG 為什麼不適合回答「有幾筆」？把計算交給程式
一般 RAG 只看取回的局部片段，不適合保證完整計數、加總與存在性判斷。本文以交易 CSV 說明如何用 deterministic code 計算，再由 LLM 解釋結果。
-
RAG 優化實戰：先建立評測，再調整檢索管線
RAG 檢索品質不穩時，先建立評測基準，再逐項檢查 embedding、chunking、reranker、OCR，以及 PDF 頁碼與 citation metadata。
-
銀行為什麼自架 LLM？從資料風險到治理要求的實務考量
自架不是銀行導入 LLM 的唯一答案。本文從資料分級、access control、vector database、PII masking 與 air-gapped environment 維運，整理架構評估方式。
-
Why RAG Can’t Answer “How Many?”: Let Deterministic Code Do the Math
Standard RAG sees retrieved fragments, not a complete dataset. Use deterministic code for counts, sums, filtering,…
-
RAG Optimization in Practice: Measure First, Then Tune Retrieval
When RAG retrieval is unstable, establish an evaluation baseline before tuning embeddings, chunking, reranking, OCR, and…
-
Why Banks Self-Host LLMs: Data Risk, Governance, and Architecture Tradeoffs
Self-hosting is not the only answer for banks adopting LLMs. Start with data classification, then evaluate…
-
ITIL 4 Exam Preparation: How I Used AI as My Tutor
How I used authorized practice material and AI-assisted error analysis to pass six advanced ITIL 4…
-
ITIL 4 and AI: Why Service Delivery Matters When Building Gets Cheap
AI lowers some development costs, but it does not supply reliability, accountability, or customer experience. ITIL…