AI工程大模型应用开发实战 (奇普·萱) (z-library.sk, 1lib.sk, z-lib.sk)

【One-Line Pitch】 A practical field manual for developers who want to build production-grade AI applications on top of existing foundation models (LLMs and LMMs), covering everything from model selection and prompt engineering to RAG, fine-tuning, and inference optimization. Read this if you are a software engineer or technical lead who needs a decision framework rather than another tutorial on calling an API. 【Book Arc】 - **Opening (~0%–10%)**: Establishes the core philosophy of "AI engineering" as distinct from model research — the goal is to build reliable, scalable applications using pre-trained foundation models, not to train models from scratch. It frames the central challenge: making sound technical choices in a fast-moving ecosystem. - **Early (~10%–30%)**: Walks through the foundational layers of the AI engineering stack, starting with how to select and evaluate the right base model for a given task. This stage introduces the key trade-offs between model capability, cost, latency, and licensing, and sets up the evaluation mindset that recurs throughout the book. - **Middle (~30%–60%)**: Dives into the core techniques for controlling model behavior without retraining: prompt engineering, Retrieval-Augmented Generation (RAG), and agent architectures. This is the practical heart of the book, showing how to ground models in external knowledge and orchestrate multi-step workflows. - **Late (~60%–85%)**: Moves into deeper customization and performance work — fine-tuning strategies, dataset engineering (curation, labeling, quality control), and inference optimization (quantization, batching, caching). The emphasis shifts from "making it work" to "making it fast, cheap, and robust." - **Ending (~85%–100%)**: Ties everything together with an AI engineering architecture view — how to design the overall system, integrate the pieces, and operate AI applications in production. The book closes with guidance on making scientific, repeatable decisions rather than chasing hype. 【Key Takeaways】 - **AI engineering is a discipline, not a wrapper** (Early): Building on foundation models requires a systematic framework for model selection, evaluation, and integration — treating LLMs as components in a larger software system rather than as magic black boxes. - **Evaluation is the foundation of every decision** (Early): You cannot improve what you cannot measure; the book stresses building task-specific evaluation sets early so that every later choice (prompt, model, fine-tune) is made against objective criteria. - **Prompt engineering is the first lever, not the only one** (Middle): Structured prompting techniques (few-shot, chain-of-thought, output formatting) can solve many problems cheaply, but they have limits — knowing when prompts are insufficient is a core skill. - **RAG grounds models in reality** (Middle): Retrieval-Augmented Generation is the primary pattern for injecting up-to-date or domain-specific knowledge into an LLM without retraining, and it requires careful attention to chunking, embedding, and retrieval quality. - **Agents extend models from chat to action** (Middle): Moving from single-turn Q&A to multi-step agent workflows (planning, tool use, memory) unlocks real automation but introduces new failure modes — error handling and observability become critical. - **Fine-tuning is a surgical tool, not a default** (Late): Fine-tuning should be reserved for changing model behavior or style, not for injecting knowledge (that is RAG's job); the book covers when to fine-tune, what data you need, and how to avoid catastrophic forgetting. - **Dataset engineering determines ceiling quality** (Late): The quality of your fine-tuning or evaluation data matters more than the quantity — curation, deduplication, and label consistency are the highest-leverage activities in any AI project. - **Inference optimization is where cost lives** (Late): Techniques like quantization, batching, and caching can cut inference costs by an order of magnitude, and the book provides practical guidance on when each technique pays off. 【Reading Tips】 - **Skim the opening philosophy** (~0–10%) if you are already convinced that AI engineering matters; the real value starts with the model selection and evaluation chapters. - **Deep-read the RAG and agent sections** (Middle) — these are the most actionable patterns for most real-world applications, and the book's examples here are the most concrete. - **Treat the fine-tuning and dataset chapters** (Late) as a reference to return to when you actually need them; the decision frameworks (when to fine-tune vs. use RAG vs. prompt) are worth memorizing even if you skip the details. - **Watch for the recurring evaluation-first mindset** — if you take away only one habit, make it building an evaluation set before you start optimizing anything. - **The final architecture chapter** (Ending) is best read as a checklist for production readiness — skim it before you deploy, not before you prototype. 【Coverage Limits】 The excerpts provided cover only the book's overall structure and table-of-contents level detail; specific examples, case studies, and code samples are not included in this guide. The guide reflects the book's stated scope and chapter progression, not its detailed technical content.