Source index and reading estimates

00-START-HERE This index links all fifty-seven content units in the Notion edition: forty-one chapters, nine front-matter units and seven appendices. Original physical-PDF spans and printed page labels are preserved as supplied metadata, not newly verified PDF pagination or live Notion positions. Reading estimates below were recomputed from the revised page-text snapshots, not inferred from PDF page count. Count word-like tokens in prose, headings, tables and link labels; exclude equations, code/diagram blocks, inline code, URLs and citation markers. At an assumed $200\,\mathrm{words/min}$, use $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}$. Unicode letter/number runs count as tokens, retaining internal apostrophes/hyphens. This is a rounded prose-only planning proxy, not measured personal reading time; mathematics, diagrams, verification, exercises and practice require additional time. Counts are snapshots, not live-updating counters.

Suggested reading order

Fast interview loop

1. Start with 00_05_for_interview_candidates and 01_what_a_rag_interview_actually_tests for the source’s practice rubric.
2. Read 03_a_repeatable_framework_for_any_rag_design_question for a useful design lens, not an exhaustive free-parameter theorem.
3. Read Chapters 15 through 18 for search, scoped memory, operations and classical IR.
4. Read Chapters 21, 22, 24, 25 and 26 for hybrid retrieval, reranking, reformulation, routing and loops.
5. Read Chapters 30 through 34 for context, support and diagnostic tests allowing multiple causes.
6. Read Chapters 35 through 38 for credibility, provenance, systems and federation.
7. Finish with 41_end_to_end_design_drills and Appendices A1, A2, A3 and A6. This broad route is distinct from the narrower one-week plan and is not a readiness guarantee.

Complete technical path

1. Read front matter units $00.01$ through $00.09$ for scope, navigation, notation, and study method.
2. Read Chapters 1 through 14 for design framing, generator behavior, prompting, chunking, parsing, and embedding.
3. Read Chapters 15 through 29 for indexes, retrieval, ranking, reformulation, control, and training data.
4. Read Chapters 30 through 41 for context, attribution, evaluation, credibility, adversarial robustness, systems, federation, multimodality, graphs, and design drills.
5. Use Appendices A1 through A7 as formula, practice, checklist, reporting, reading, notation, and glossary references.

Production system path

1. Begin with Chapters 3, 12, 13 and 14 for framing, representation, retrieval units and parsing.
2. Continue with Chapters 15, 16, 17, 21 and 22 for capacity and serving trade-offs.
3. Add Chapters 24, 25, 26 and 30 for query/control and context decisions.
4. Validate with Chapters 31 through 38 and A4_the_rag_card_and_index_datasheet.
5. Add Chapters 39 and 40 when modality, graph structure, provenance or global-query requirements justify them.
6. Use 41_end_to_end_design_drills and A3_design_checklists for practice and review, then complete workload-specific tests and release requirements. Reading the route does not itself approve deployment.

Evaluation and trust path

1. Read Chapters 5, 10 and 11 for data/legal boundaries, prompt sensitivity, abstention and calibration.
2. Read Chapters 27 through 29 for generator/retriever training and synthetic-data limitations.
3. Read Chapters 31 through 36 for citation support, retrieval/generation metrics, end-to-end tests, credibility, provenance and attacks.
4. Read Chapters 37 and 38 for operational work, cost and distributed governance.
5. Complete A4_the_rag_card_and_index_datasheet using measured artifacts, explicit uncertainty and domain-specific review; this is not a universal compliance certification.

Unit index

00.01. Title Page

Chapter page: 00_01_title_page
Summary: Preserve the source title, named author/institute, edition label and contact channels as attributed metadata, without independently certifying credentials or a current-version guarantee.[1]
Source span: Physical PDF pages 1 to 1. Printed pages unnumbered.
Prose-only planning estimate: $W=1{,}280$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=7\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[1]

00.02. Contents

Chapter page: 00_02_contents
Summary: This unit turns printed contents pages i through x into an interview-oriented navigation map for Retrieval-Augmented Generation (RAG).
Source span: Physical PDF pages 2 to 11. Printed pages i to x.
Prose-only planning estimate: $W=3{,}482$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=18\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[2]

00.03. Preface

Chapter page: 00_03_preface
Summary: This preface explains how the book turns Retrieval-Augmented Generation (RAG) from a component list into a production reasoning discipline for artificial intelligence systems.
Source span: Physical PDF pages 12 to 13. Printed pages xi to xii.
Prose-only planning estimate: $W=4{,}487$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=23\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[3]

00.04. How to Use This Book

Chapter page: 00_04_how_to_use_this_book
Summary: This unit explains how to choose a reading route, use the book’s recurring section structure, interpret its interview tags and references, and turn its appendices into preparation tools.
Source span: Physical PDF pages 14 to 18. Printed pages xiii to xvii.
Prose-only planning estimate: $W=4{,}484$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=23\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[4]

00.05. For Interview Candidates

Chapter page: 00_05_for_interview_candidates
Summary: This guide turns Retrieval-Augmented Generation (RAG) interview preparation into a decision-first plan for design, derivation, debugging, pacing, and practice.
Source span: Physical PDF pages 19 to 23. Printed pages xviii to xxii.
Prose-only planning estimate: $W=6{,}325$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=32\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[5]

00.06. For Practicing Engineers

Chapter page: 00_06_for_practicing_engineers
Summary: This unit maps the technical content of Retrieval-Augmented Generation (RAG) to the decisions, failures, costs, and trade-offs that practicing engineers face.
Source span: Physical PDF pages 24 to 27. Printed pages xxiii to xxvi.
Prose-only planning estimate: $W=4{,}474$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=23\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[6]

00.07. Notation and Symbols

Chapter page: 00_07_notation_and_symbols
Summary: Clarify local symbol definitions, overloads, units, precision and measurement conventions rather than forcing every chapter to use one global meaning for each symbol.[7]
Source span: Physical PDF pages 28 to 31. Printed pages xxvii to xxx.
Prose-only planning estimate: $W=2{,}819$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=15\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[7]

00.08. Acknowledgments

Chapter page: 00_08_acknowledgments
Summary: Explain who and what made the book possible, how those sources shaped its practical and evidence-based character, and why negative results matter to its honesty.
Source span: Physical PDF pages 32 to 32. Printed pages xxxi.
Prose-only planning estimate: $W=2{,}525$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=13\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[8]

00.09. About the Author

Chapter page: 00_09_about_the_author
Summary: Preserve the supplied author profile, original-authorship statement and correction routes while attributing biography and undated audience counts rather than treating them as independently verified live facts.[9]
Source span: Physical PDF pages 33 to 33. Printed pages xxxii.
Prose-only planning estimate: $W=2{,}100$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=11\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[9]

Chapter 1. What a RAG Interview Actually Tests

Chapter page: 01_what_a_rag_interview_actually_tests
Summary: Use the source’s interview lens to discuss knowledge location, explicit costs and conditions under which retrieval helps or harms; this is a practice framework, not an industry-wide hiring-outcome study.[10]
Source span: Physical PDF pages 35 to 54. Printed pages 2 to 21.
Prose-only planning estimate: $W=11{,}138$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=56\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[10]

Chapter 2. The Competing Answers to the Same Problem

Chapter page: 02_the_competing_answers_to_the_same_problem
Summary: Choose among continual learning, model editing, long context, generative retrieval, and retrieval-augmented generation by matching each method to the constraint and cost shape it actually serves.
Source span: Physical PDF pages 55 to 75. Printed pages 22 to 42.
Prose-only planning estimate: $W=12{,}803$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=65\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[11]

Chapter 3. A Repeatable Framework for Any RAG Design Question

Chapter page: 03_a_repeatable_framework_for_any_rag_design_question
Summary: This chapter gives one repeatable way to turn an underspecified retrieval design prompt into explicit choices, priced trade-offs, and a defensible interview answer.
Source span: Physical PDF pages 76 to 95. Printed pages 43 to 62.
Prose-only planning estimate: $W=10{,}592$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=53\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[12]

Chapter 4. What the LLM Brings and What It Breaks

Chapter page: 04_what_the_llm_brings_and_what_it_breaks
Summary: Explain model contributions and failure mechanisms, then test plausible causes and interactions before assigning corrective work.[13]
Source span: Physical PDF pages 97 to 121. Printed pages 64 to 88.
Prose-only planning estimate: $W=9{,}639$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=49\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[13]

Chapter 5. Data, Privacy, and the Legal Surface

Chapter page: 05_data_privacy_and_the_legal_surface
Summary: This chapter prepares you to explain how Retrieval-Augmented Generation (RAG) architecture changes the cost, evidence, and reversibility of data decisions without claiming that it resolves the underlying legal questions.
Source span: Physical PDF pages 122 to 144. Printed pages 89 to 111.
Prose-only planning estimate: $W=10{,}830$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=55\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[14]

Chapter 6. Scaling Laws and the Economics of Retrieval vs Parameters

Chapter page: 06_scaling_laws_and_the_economics_of_retrieval_vs_parameters
Summary: This chapter explains how to allocate training, serving, memory, and retrieval budgets for a Retrieval-Augmented Generation (RAG) system.
Source span: Physical PDF pages 145 to 167. Printed pages 112 to 134.
Prose-only planning estimate: $W=7{,}588$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=38\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[15]

Chapter 7. In-Context Learning - The Mechanism RAG Rides On

Chapter page: 07_in_context_learning_the_mechanism_rag_rides_on
Summary: This chapter explains how Retrieval-Augmented Generation (RAG) uses in-context learning (ICL), when supervised fine-tuning (SFT) is the better tool, and how selection, format, order, inference theory, and cost shape the design.
Source span: Physical PDF pages 168 to 193. Printed pages 135 to 160.
Prose-only planning estimate: $W=8{,}224$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=42\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[16]

Chapter 8. Reading the Machine: Circuits, Induction Heads, and Attribution

Chapter page: 08_reading_the_machine_circuits_induction_heads_and_attribution
Summary: Distinguish evidence about one model run from evidence about fixed model weights, then use residual streams, query-key circuits, output-value circuits, and induction heads to reason about Retrieval-Augmented Generation (RAG) failures without claiming more than the evidence supports.
Source span: Physical PDF pages 194 to 216. Printed pages 161 to 183.
Prose-only planning estimate: $W=7{,}628$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=39\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[17]

Chapter 9. Prompting for Retrieval-Augmented Systems

Chapter page: 09_prompting_for_retrieval_augmented_systems
Summary: Explain where prompting ends, where retrieval begins, how reasoning scaffolds change decoding, and why more calls or more passages can reduce answer quality.
Source span: Physical PDF pages 218 to 235. Printed pages 185 to 202.
Prose-only planning estimate: $W=7{,}917$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=40\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[18]

Chapter 10. Prompt Sensitivity

Chapter page: 10_prompt_sensitivity
Summary: This chapter explains how meaning-preserving prompt changes can alter Retrieval-Augmented Generation (RAG), how to measure the effect, and how to make defensible template and optimization decisions.
Source span: Physical PDF pages 236 to 256. Printed pages 203 to 223.
Prose-only planning estimate: $W=9{,}418$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=48\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[19]

Chapter 11. Abstention and Calibration

Chapter page: 11_abstention_and_calibration
Summary: This chapter explains when a Retrieval-Augmented Generation (RAG) system should refuse, how to build a useful confidence signal, and how to keep that signal honest for short and long answers.
Source span: Physical PDF pages 257 to 283. Printed pages 224 to 250.
Prose-only planning estimate: $W=8{,}802$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=45\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[20]

Chapter 12. Text Representation

Chapter page: 12_text_representation
Summary: This chapter explains how text representations trade exact matching, semantic reach, operational stability, training cost, and privacy in a Retrieval-Augmented Generation (RAG) system.
Source span: Physical PDF pages 285 to 306. Printed pages 252 to 273.
Prose-only planning estimate: $W=9{,}781$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=49\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[21]

Chapter 13. Chunking and Granularity

Chapter page: 13_chunking_and_granularity
Summary: This chapter explains how to choose retrieval units, route across granularities, preserve enough context, and measure whether a chunker actually improved retrieval.
Source span: Physical PDF pages 307 to 332. Printed pages 274 to 299.
Prose-only planning estimate: $W=10{,}106$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=51\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[22]

Chapter 14. Beyond Plain Text - Tables, Layout, Documents

Chapter page: 14_beyond_plain_text_tables_layout_documents
Summary: This chapter explains why table cells and document blocks lose meaning when they are flattened, how table and layout encoders restore structure, and which costs, invariance limits, metrics, and layout priors matter in a Retrieval-Augmented Generation (RAG) interview.
Source span: Physical PDF pages 333 to 355. Printed pages 300 to 322.
Prose-only planning estimate: $W=9{,}541$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=48\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[23]

Chapter 15. Approximate Nearest Neighbor Search

Chapter page: 15_approximate_nearest_neighbor_search
Summary: This chapter explains how to choose and defend an Approximate Nearest Neighbor (ANN) index for Retrieval-Augmented Generation (RAG) under recall, latency, memory, build, and update constraints.
Source span: Physical PDF pages 357 to 383. Printed pages 324 to 350.
Prose-only planning estimate: $W=10{,}797$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=54\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[24]

Chapter 16. Compression and Index Economics

Chapter page: 16_compression_and_index_economics
Summary: Build a defensible retrieval index budget from code length, scoring error, scan shape, and fleet multipliers without confusing a smaller stored vector with a smaller system.
Source span: Physical PDF pages 384 to 406. Printed pages 351 to 373.
Prose-only planning estimate: $W=8{,}270$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=42\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[25]

Chapter 17. Operating a Vector Store

Chapter page: 17_operating_a_vector_store
Summary: This chapter prepares you to operate the retrieval layer of a Retrieval-Augmented Generation (RAG) system under deletion, growth, metadata, filtering, and natural-language constraints.
Source span: Physical PDF pages 407 to 434. Printed pages 374 to 401.
Prose-only planning estimate: $W=12{,}994$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=65\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[26]

Chapter 18. Classical IR You Are Expected to Know

Chapter page: 18_classical_ir_you_are_expected_to_know
Summary: This chapter prepares you to explain how classical lexical ranking supports a Retrieval-Augmented Generation (RAG) system, where its assumptions fail, and why sparse retrieval still protects rare queries.
Source span: Physical PDF pages 436 to 459. Printed pages 403 to 426.
Prose-only planning estimate: $W=10{,}185$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=51\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[27]

Chapter 19. Learning to Rank

Chapter page: 19_learning_to_rank
Summary: Compare metric-aware ranking objectives and differentiable surrogates, sign conventions and serving architectures without treating every surrogate gradient as an exact gradient of a discontinuous ranking metric.[28]
Source span: Physical PDF pages 460 to 481. Printed pages 427 to 448.
Prose-only planning estimate: $W=11{,}731$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=59\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[28]

Chapter 20. Dense Retrieval

Chapter page: 20_dense_retrieval
Summary: This chapter prepares a Retrieval-Augmented Generation (RAG) candidate to explain dense retrieval, train it, price it, combine it with lexical search, and extend it across languages.
Source span: Physical PDF pages 482 to 504. Printed pages 449 to 471.
Prose-only planning estimate: $W=14{,}208$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=72\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[29]

Chapter 21. Learned Sparse and Multi-Vector Retrieval

Chapter page: 21_learned_sparse_and_multi_vector_retrieval
Summary: This chapter prepares you to explain how learned lexical weights, token-level matching, compression, pruning, and rank fusion trade retrieval quality against index size and latency.
Source span: Physical PDF pages 505 to 531. Printed pages 472 to 498.
Prose-only planning estimate: $W=12{,}925$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=65\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[30]

Chapter 22. Reranking

Chapter page: 22_reranking
Summary: Decide when a second ranking stage is worth its cost, how to size candidates and evidence separately, and how to evaluate cross-encoders and LLM ranking modes under matched constraints.[31]
Source span: Physical PDF pages 532 to 562. Printed pages 499 to 529.
Prose-only planning estimate: $W=9{,}649$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=49\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[31]

Chapter 23. Generative Retrieval

Chapter page: 23_generative_retrieval
Summary: Compare document-identifier generation with external indexing, including update/deletion, serving and evidence-access trade-offs, without assuming that every generative retrieval design has identical update costs.[32]
Source span: Physical PDF pages 563 to 580. Printed pages 530 to 547.
Prose-only planning estimate: $W=10{,}967$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=55\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[32]

Chapter 24. Query Reformulation

Chapter page: 24_query_reformulation
Summary: This chapter prepares you to explain how Retrieval-Augmented Generation (RAG) systems diagnose weak queries, expand or rewrite them, train retrieval-aware representations, repair conversational context, and search with hypothetical answers.
Source span: Physical PDF pages 582 to 607. Printed pages 549 to 574.
Prose-only planning estimate: $W=12{,}387$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=62\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[33]

Chapter 25. Routing and Adaptive Retrieval

Chapter page: 25_routing_and_adaptive_retrieval
Summary: This chapter explains how a Retrieval-Augmented Generation (RAG) system decides when to retrieve, where to look, which operation to run, and how to fail safely on unfamiliar queries.
Source span: Physical PDF pages 608 to 636. Printed pages 575 to 603.
Prose-only planning estimate: $W=13{,}657$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=69\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[34]

Chapter 26. Iterative, Recursive, and Agentic Retrieval

Chapter page: 26_iterative_recursive_and_agentic_retrieval
Summary: Compare when iterative, recursive or agentic retrieval adds value, how to form later queries, and how to stop when extra calls or fan-out no longer justify their cost and risk.[35]
Source span: Physical PDF pages 637 to 674. Printed pages 604 to 641.
Prose-only planning estimate: $W=11{,}011$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=56\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[35]

Chapter 27. Fine-tuning the Generator

Chapter page: 27_fine_tuning_the_generator
Summary: Evaluate generator-side training interventions under observed retrieval defects, with explicit objectives and budgets, while distinguishing behavioral change from knowledge deletion, invariance and complete-system memory savings.[36]
Source span: Physical PDF pages 676 to 696. Printed pages 643 to 663.
Prose-only planning estimate: $W=9{,}549$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=48\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[36]

Chapter 28. Training the Retriever

Chapter page: 28_training_the_retriever
Summary: decide which Retrieval-Augmented Generation (RAG) component to train, connect generator utility to a discrete retriever, and price the index work that retriever updates create.
Source span: Physical PDF pages 697 to 727. Printed pages 664 to 694.
Prose-only planning estimate: $W=9{,}696$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=49\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[37]

Chapter 29. Bootstrapping Training Data

Chapter page: 29_bootstrapping_training_data
Summary: Manufacture and filter retrieval supervision while preserving the target relevance relation, measuring distribution shift and distinguishing teacher-signal limits from a universal student-quality ceiling.[38]
Source span: Physical PDF pages 728 to 759. Printed pages 695 to 726.
Prose-only planning estimate: $W=9{,}500$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=48\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[38]

Chapter 30. Lost in the Middle

Chapter page: 30_lost_in_the_middle
Summary: This chapter prepares you to diagnose and mitigate position-dependent failures in Retrieval-Augmented Generation (RAG) systems.
Source span: Physical PDF pages 761 to 790. Printed pages 728 to 757.
Prose-only planning estimate: $W=9{,}347$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=47\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[39]

Chapter 31. Attribution and Citation

Chapter page: 31_attribution_and_citation
Summary: Distinguish citation presence, evidential support, factual truth and actual model use, then define and evaluate the claimed form of attribution.[40]
Source span: Physical PDF pages 791 to 809. Printed pages 758 to 776.
Prose-only planning estimate: $W=8{,}361$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=42\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[40]

Chapter 32. Evaluating Retrieval

Chapter page: 32_evaluating_retrieval
Summary: Evaluate coverage and ordering under declared reference sets, eligibility and aggregation; distinguish measured relevance recall from judge-based or label-free proxies when gold judgments are missing.[41]
Source span: Physical PDF pages 811 to 825. Printed pages 778 to 792.
Prose-only planning estimate: $W=8{,}527$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=43\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[41]

Chapter 33. Evaluating Generation

Chapter page: 33_evaluating_generation
Summary: This chapter explains how to choose and interpret generation metrics for a Retrieval-Augmented Generation (RAG) system when oracle access, claim coverage, output length, and summarization behavior all affect the result.
Source span: Physical PDF pages 826 to 847. Printed pages 793 to 814.
Prose-only planning estimate: $W=8{,}894$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=45\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[42]

Chapter 34. Evaluating the System

Chapter page: 34_evaluating_the_system
Summary: This chapter is for diagnosing, validating, and stress-testing a Retrieval-Augmented Generation (RAG) system before a launch decision.
Source span: Physical PDF pages 848 to 870. Printed pages 815 to 837.
Prose-only planning estimate: $W=9{,}043$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=46\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[43]

Chapter 35. Source Credibility

Chapter page: 35_source_credibility
Summary: Measure defined credibility signals and source-policy outcomes while separating them from truth guarantees, and reason explicitly about freshness, pluralism and exposure fairness.[44]
Source span: Physical PDF pages 872 to 902. Printed pages 839 to 869.
Prose-only planning estimate: $W=10{,}312$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=52\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[44]

Chapter 36. Provenance and Adversarial Robustness

Chapter page: 36_provenance_and_adversarial_robustness
Summary: Evaluate provenance authentication, external-content attacks and access boundaries without treating authentic provenance as proof of truth or a defense as a universal safety guarantee.[45]
Source span: Physical PDF pages 903 to 933. Printed pages 870 to 900.
Prose-only planning estimate: $W=10{,}345$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=52\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[45]

Chapter 37. Latency, Cost, and Systems

Chapter page: 37_latency_cost_and_systems
Summary: Build a stage-level model of Retrieval-Augmented Generation (RAG) latency and cost, then use caching, batching, kernel design, pipelining, retrieval cascades, and multi-resolution indexing only where the measured bottleneck justifies them.
Source span: Physical PDF pages 935 to 965. Printed pages 902 to 932.
Prose-only planning estimate: $W=11{,}956$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=60\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[46]

Chapter 38. Distributed and Federated RAG

Chapter page: 38_distributed_and_federated_rag
Summary: This chapter is for designing Retrieval-Augmented Generation (RAG) across independently governed sources, regions, tenants, and shared public infrastructure.
Source span: Physical PDF pages 966 to 984. Printed pages 933 to 951.
Prose-only planning estimate: $W=9{,}240$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=47\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[47]

Chapter 39. Multimodal RAG

Chapter page: 39_multimodal_rag
Summary: This chapter explains how Retrieval-Augmented Generation (RAG) systems index, retrieve, fuse, generate from, attribute, and evaluate text, images, tables, audio, and video.
Source span: Physical PDF pages 985 to 1019. Printed pages 952 to 986.
Prose-only planning estimate: $W=10{,}183$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=51\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[48]

Chapter 40. Graph RAG

Chapter page: 40_graph_rag
Summary: This chapter prepares you to defend every stage, cost, failure, and claim boundary of graph-based Retrieval-Augmented Generation (RAG), from construction and grounded retrieval through generation and the GraphRAG versus LightRAG choice.
Source span: Physical PDF pages 1020 to 1054. Printed pages 987 to 1021.
Prose-only planning estimate: $W=10{,}295$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=52\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[49]

Chapter 41. End-to-End Design Drills

Chapter page: 41_end_to_end_design_drills
Summary: Turn seven Retrieval-Augmented Generation (RAG) interview prompts into measurable architectures, explicit trade-offs, and answers that start from the binding constraint rather than a product name.
Source span: Physical PDF pages 1055 to 1082. Printed pages 1022 to 1049.
Prose-only planning estimate: $W=10{,}269$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=52\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[50]

Appendix A1. Formula Sheet

Chapter page: A1_formula_sheet
Summary: Collect eight formula groups with explicit definitions, domains, normalization and approximation boundaries: BM25, RRF, MRR, nDCG, modularity, PQ, compute/latency and index-memory arithmetic.[51]
Source span: Physical PDF pages 1083 to 1092. Printed pages 1050 to 1059.
Prose-only planning estimate: $W=5{,}320$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=27\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[51]

Appendix A2. Question Bank

Chapter page: A2_question_bank
Summary: This appendix turns the book’s Retrieval-Augmented Generation (RAG) material into a timed self-test with Core, Senior, and Staff prompts plus compact answer spines.
Source span: Physical PDF pages 1093 to 1110. Printed pages 1060 to 1077.
Prose-only planning estimate: $W=15{,}992$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=80\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[52]

Appendix A3. Design Checklists

Chapter page: A3_design_checklists
Summary: Use these five operational checklists to turn a Retrieval-Augmented Generation (RAG) design review into named decisions, measured quantities, and explicit exit criteria.
Source span: Physical PDF pages 1111 to 1116. Printed pages 1078 to 1083.
Prose-only planning estimate: $W=7{,}737$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=39\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[53]

Appendix A4. The RAG Card and Index Datasheet

Chapter page: A4_the_rag_card_and_index_datasheet
Summary: This appendix supplies two handoff templates for a Retrieval-Augmented Generation (RAG) system, with the RAG Card documenting scope, retrieval decisions, measured behavior, and known failures while the Index Datasheet documents artifact composition, representation, structure, footprint, performance, and lifecycle.
Source span: Physical PDF pages 1117 to 1121. Printed pages 1084 to 1088.
Prose-only planning estimate: $W=5{,}416$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=28\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[54]

Appendix A5. Annotated Reading List

Chapter page: A5_annotated_reading_list
Summary: Use the chapter-mapped reading list’s 161 annotated entries (160 distinct titles, with one repeated title) and 52 core selections to locate assumptions and results; working through arithmetic is not independent replication of every empirical study.[55]
Source span: Physical PDF pages 1122 to 1133. Printed pages 1089 to 1100.
Prose-only planning estimate: $W=11{,}306$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=57\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[55]

Appendix A6. Notation Quick Reference

Chapter page: A6_notation_quick_reference
Summary: This appendix provides a compact, calculation-oriented reference for the source’s symbols, overload rules, and conventions while leaving the extended discussion to the front matter.
Source span: Physical PDF pages 1134 to 1135. Printed pages 1101 to 1102.
Prose-only planning estimate: $W=1{,}972$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=10\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[56]

Appendix A7. Glossary

Chapter page: A7_glossary
Summary: Resolve overlapping vocabulary and local notation while keeping payloads, measured systems, guarantees, heuristics and empirical claims distinct.[57]
Source span: Physical PDF pages 1136 to 1141. Printed pages 1103 to 1108.
Prose-only planning estimate: $W=6{,}468$ word-like tokens; $t_{\mathrm{prose}}=\lceil W/200\rceil\,\mathrm{min}=33\,\mathrm{min}$ at assumed $200\,\mathrm{words/min}$.[57]