大模型、多模态、推理模型和 AI 产品趋势
This collection features articles and community discussions in the AI 前沿 space, focusing on real-world engineering scenarios, tool selection, practical paths, and risk boundaries.
The page is continuously updated and serves as a learning roadmap, technology selection reference, and technical review entry point.
LLM applications need SLIs and SLOs, but traditional Web targets do not transfer: latency varies widely, quality is hard to quantify, and cost is itself a metric. This article covers latency (p95/TTFT), availability (model API failures), quality (evaluation scores/bad-case rate), cost (tokens/request), alerting, and error budgets.
Production RAG almost always needs conditional retrieval: search only a department, time window, or product. Vector recall alone cannot filter precisely. This article explains metadata design for chunks, pgvector SQL and Qdrant payload filters, pre-filter performance benefits, and permission isolation.
When RAG answers poorly, retrieval is usually at fault--but first determine whether the issue is missed recall or irrelevant retrieval. This article presents a systematic workflow for inspecting recall, similarity, reranking, and prompts, plus ways to diagnose bad chunking, embedding mismatches, and overly strict filters.