Appendix G — Appendix G: Further reading

Pointers for going deeper, organized by where they fit in the book. Links and exact titles may drift, so search for the current version.

G.1 Foundational essays

Anthropic, “Building effective agents” (2024). The workflows-versus-agents framing we use in Part IV, with the five workflow patterns. The most useful brief introduction for anyone building agentic systems.
Anthropic, “Introducing Contextual Retrieval” (2024). The technique at the center of Chapter 8, with the numbers on how much it reduces failed retrievals.
Anthropic, “Introducing the Model Context Protocol” (2024). The standard behind Chapter 12; the spec and SDKs for writing your own tools.

G.2 Retrieval and RAG

Documentation for ChromaDB and other vector databases, for when the in-memory store outgrows a laptop.
The BGE-M3 and BGE reranker model cards (BAAI), for the embedding and reranking models used throughout Part II and III.
Surveys of RAG architectures (naive, advanced, agentic) for the techniques beyond this book’s core.

G.3 Evaluation and observability

RAGAS, DeepEval, and Braintrust documentation, for evaluation frameworks that formalize Chapter 9’s toolkit.
Langfuse, Arize Phoenix, and LangSmith, for the tracing and production-evaluation tooling of Chapter 24. Langfuse is open-source and self-hostable, which suits privacy-conscious deployments.

G.4 Fine-tuning and synthetic data

The PEFT, TRL, and bitsandbytes library documentation, for the LoRA and QLoRA mechanics of Chapter 16.
Research on model collapse and the real-plus-synthetic mix, for the deeper story behind Chapter 17’s rule.

G.5 Multimodal and time series

Whisper (OpenAI) and pyannote.audio for speech (Chapter 18).
Chronos (Amazon) and other time-series foundation models for Chapter 20.

G.6 Safety and compliance

The OWASP Top 10 for LLM Applications, the canonical list behind Chapter 23, updated as new attack classes emerge.
The NIST AI Risk Management Framework and the EU AI Act text, for the governance side of Appendix D. For US state law, go to the current statute text, because secondary summaries age quickly.

G.7 On the craft

The best further reading is your own evaluation set. Because every technique in this book is worth exactly as much as it scores on your data, the habit of measuring is what will keep you current as the models and tools change, long after the list above has aged.

--- title: "Appendix G: Further reading" --- Pointers for going deeper, organized by where they fit in the book. Links and exact titles may drift, so search for the current version. ## Foundational essays - **Anthropic, "Building effective agents" (2024).** The workflows-versus-agents framing we use in Part IV, with the five workflow patterns. The most useful brief introduction for anyone building agentic systems. - **Anthropic, "Introducing Contextual Retrieval" (2024).** The technique at the center of Chapter 8, with the numbers on how much it reduces failed retrievals. - **Anthropic, "Introducing the Model Context Protocol" (2024).** The standard behind Chapter 12; the spec and SDKs for writing your own tools. ## Retrieval and RAG - Documentation for **ChromaDB** and other vector databases, for when the in-memory store outgrows a laptop. - The **BGE-M3** and **BGE reranker** model cards (BAAI), for the embedding and reranking models used throughout Part II and III. - Surveys of RAG architectures (naive, advanced, agentic) for the techniques beyond this book's core. ## Evaluation and observability - **RAGAS**, **DeepEval**, and **Braintrust** documentation, for evaluation frameworks that formalize Chapter 9's toolkit. - **Langfuse**, **Arize Phoenix**, and **LangSmith**, for the tracing and production-evaluation tooling of Chapter 24. Langfuse is open-source and self-hostable, which suits privacy-conscious deployments. ## Fine-tuning and synthetic data - The **PEFT**, **TRL**, and **bitsandbytes** library documentation, for the LoRA and QLoRA mechanics of Chapter 16. - Research on **model collapse** and the real-plus-synthetic mix, for the deeper story behind Chapter 17's rule. ## Multimodal and time series - **Whisper** (OpenAI) and **pyannote.audio** for speech (Chapter 18). - **Chronos** (Amazon) and other time-series foundation models for Chapter 20. ## Safety and compliance - The **OWASP Top 10 for LLM Applications**, the canonical list behind Chapter 23, updated as new attack classes emerge. - The **NIST AI Risk Management Framework** and the **EU AI Act** text, for the governance side of Appendix D. For US state law, go to the current statute text, because secondary summaries age quickly. ## On the craft The best further reading is your own evaluation set. Because every technique in this book is worth exactly as much as it scores on your data, the habit of measuring is what will keep you current as the models and tools change, long after the list above has aged.