00.02 Contents

00-START-HERE · 01-READING-ROUTES · 00_01_title_page · 00_03_preface This unit maps printed contents pages i through x to the book’s topics. Printed page numbers are source-edition references, not live Notion positions.

TL;DR

The book moves from interview framing and generator limits through prompting, representation, indexing, retrieval, routing, training, context assembly, evaluation, trust, scaling, and design drills.
Parts I through III establish the problem, the generator, and the context interface.
Parts IV through VII trace the retrieval path from representation to indexing, ranking, reformulation, routing, and multi-step control flow.
Parts VIII through XII cover training, assembly, evaluation, credibility, robustness, systems, advanced variants, and full design exercises.
Appendices A through G provide formulas, question banks, design checklists, documentation templates, readings, notation, and terminology.

The story

Treat the contents as a metro map for a RAG interview. The front matter is the ticket hall. It explains who the route serves, how to read the signs, and which notation appears along the way. Part I is the central interchange because it frames the interview, compares competing answers, and supplies a repeatable design route. The next lines travel through the system itself. The generator, prompting, representation, indexing, retrieval, query control, training, and context assembly each get their own line. Every chapter is a station where an interviewer can stop the conversation and ask for a mechanism, trade-off, failure mode, or design choice. The final lines are the inspection route. Evaluation checks retrieval, generation, and the complete system. Trust and robustness inspect the evidence and threat surface. Scaling and advanced variants test whether the route still works under operational pressure. The appendices are the service depot, holding formulas, drills, checklists, documentation, readings, notation, and a glossary for rapid repair before an interview.

Decoder table

Map line	Printed-page start	Interview use
Part I. The RAG Interview Landscape	1	Frame the problem, compare alternatives, and structure a design answer.
Part II. The Generator Side	63	Explain what the generator contributes and where it fails.
Part III. Prompting and Context Construction	184	Control instructions, reasoning scaffolds, sensitivity, abstention, and calibration.
Part IV. Representing What You Retrieve	251	Choose representations, chunking granularity, and document structure.
Part V. Indexing and Vector Search	323	Connect search structures, compression, and vector-store operations.
Part VI. Retrieval and Ranking	402	Compare classical, dense, sparse, multi-vector, reranking, and generative retrieval.
Part VII. Query Understanding and Control Flow	548	Reformulate, route, iterate, recurse, and stop retrieval.
Part VIII. Training the RAG System	642	Train generators and retrievers, then bootstrap training data.
Part IX. Generation and Context Assembly	727	Place evidence, manage positional effects, and support attribution.
Part X. Evaluation	777	Separate retrieval, generation, and system-level measurement.
Part XI. Trust, Credibility, and Adversarial Robustness	838	Examine source credibility, provenance, attacks, and privacy.
Part XII. Scaling, Advanced Variants, and Design Drills	901	Reason about systems, federation, modalities, graphs, and end-to-end cases.
Appendices A through G	1050	Rehearse formulas, questions, checklists, documentation, readings, notation, and terms.

Core mechanics

The catalog preserves source titles, section labels and printed page references. These are original-edition navigation labels, not current Notion page positions or a claim that every revised heading retains identical wording. Standard abbreviations may be explained in prose without changing quoted titles.

Front matter

Unit	Printed page
Preface	xi
How to Use This Book	xiii
For Interview Candidates	xviii
For Practicing Engineers	xxiii
Notation and Symbols	xxvii
Acknowledgments	xxxi
About the Author	xxxii

Part I. The RAG Interview Landscape

Unit	Printed page
Part I. The RAG Interview Landscape	1
Chapter 1. What a RAG Interview Actually Tests	2
$1.1$ Parametric, non-parametric, semi-parametric: the frame behind every question	2
$1.2$ The five bottlenecks that created RAG	6
$1.3$ The five benefits, and what each one costs you	10
$1.4$ RAG is one instance of tool use, not the general case	13
$1.5$ When not to retrieve: the popularity curve and retrieval-induced harm	17
Chapter 2. The Competing Answers to the Same Problem	22
$2.1$ Continual learning and catastrophic forgetting	22
$2.2$ Model editing: gradient-based and locate-and-update	26
$2.3$ Long context vs retrieval	31
$2.4$ Generative retrieval: putting the index back in the weights	35
$2.5$ Choosing between them out loud in an interview	39
Chapter 3. A Repeatable Framework for Any RAG Design Question	43
$3.1$ The five retrieval decisions as an answer skeleton	43
$3.2$ Naive -> Advanced -> Modular: naming your design’s maturity	47
$3.3$ The three injection interfaces: text, embeddings, parameters	50
$3.4$ Sizing the problem before designing it	54
$3.5$ “One size does not fit all”: adaptivity as the default answer	58

Part II. The Generator Side

Unit	Printed page
Part II. The Generator Side	63
Chapter 4. What the LLM Brings and What It Breaks	64
$4.1$ Pre-training, SFT, post-training - which stage owns which failure	64
$4.2$ Memorization: three laws and their consequences	68
$4.3$ Long-tail knowledge and co-occurrence bias	71
$4.4$ Hallucination taxonomy: factuality vs faithfulness, intrinsic vs extrinsic	76
$4.5$ The five causes of factual error	80
$4.6$ Why verification is hard: black box, moving truth, biased verifiers	84
Chapter 5. Data, Privacy, and the Legal Surface	89
$5.1$ What is actually in the pre-training corpus	89
$5.2$ PII extraction and the divergence attack	93
$5.3$ Copyright, Books3, and verbatim continuation	97
$5.4$ Unlearning and why it cannot scale	101
$5.5$ SILO: low-risk weights, high-risk datastore	106
Chapter 6. Scaling Laws and the Economics of Retrieval vs Parameters	112
$6.1$ Kaplan, Chinchilla, and post-Chinchilla token/parameter ratios	112
$6.2$ FLOPs ~= 6ND and the memory arithmetic you must do live	117
$6.3$ Three observations that change decisions: shape, embeddings, convergence	121
$6.4$ Test-time compute vs pre-training compute	126
$6.5$ The small-model-plus-datastore argument, and its limit	130
Chapter 7. In-Context Learning - The Mechanism RAG Rides On	135
$7.1$ ICL vs SFT: what each actually buys you	135
$7.2$ Demonstration selection and the test-input conjecture	140
$7.3$ Format sensitivity, and why there is no universally best format	144
$7.4$ Ordering, K! search space, and content-free calibration	148
$7.5$ Why ICL works, theory 1: Bayesian / kernel regression	152
$7.6$ Why ICL works, theory 2: gradient descent as inference	156
Chapter 8. Reading the Machine: Circuits, Induction Heads, and Attribution	161
$8.1$ Feature attribution vs mechanistic interpretability	161
$8.2$ The residual stream as an information bus	165
$8.3$ QK and OV circuits: where to look, what to move	170
$8.4$ Induction heads and “advanced copying”	175
$8.5$ What circuits let you claim about a RAG answer - and what they don’t	179

Part III. Prompting and Context Construction

Unit	Printed page
Part III. Prompting and Context Construction	184
Chapter 9. Prompting for Retrieval-Augmented Systems	185
$9.1$ The prompt hierarchy: completion -> instruction -> ICL -> RAG	185
$9.2$ Reasoning scaffolds: CoT, self-consistency, Tree and Graph of Thoughts	189
$9.3$ When more scaffolding makes things worse	194
$9.4$ Retrieving more can hallucinate more	198
Chapter 10. Prompt Sensitivity	203
$10.1$ How much formatting alone moves accuracy	203
$10.2$ Measuring sensitivity: POSIX, PSS, decoding confidence	207
$10.3$ Sensitivity as a label-free performance predictor	211
$10.4$ Few-shot beats scale for reducing sensitivity	214
$10.5$ Prompt optimization: COPLE and P2G	219
Chapter 11. Abstention and Calibration	224
$11.1$ Should a RAG system ever refuse?	224
$11.2$ Five abstention mechanisms and how each fails	228
$11.3$ Self-reflection compounds - the model-collapse connection	233
$11.4$ Calibration vs failure prediction	237
$11.5$ Eliciting confidence from a black box	242
$11.6$ Long-form calibration: $80\%$ of what?	246

Part IV. Representing What You Retrieve

Unit	Printed page
Part IV. Representing What You Retrieve	251
Chapter 12. Text Representation	252
$12.1$ Four eras: bag-of-words -> static -> contextual -> universal	252
$12.2$ Distributional semantics and meaning drift	256
$12.3$ Sparse vs dense: dimensionality, interpretability, exact match	260
$12.4$ Sentence embeddings, and why vision augmentation fails on text	265
$12.5$ What a sentence embedding actually contains	269
Chapter 13. Chunking and Granularity	274
$13.1$ Five levels of splitting	274
$13.2$ Chunking by intent: pseudo-instructions	278
$13.3$ Mixture of Granularity: routing over chunk sizes	282
$13.4$ Propositions and fine-grained retrieval	287
$13.5$ Evaluating a chunker: IoU and the segmentation metrics	291
$13.6$ Does semantic chunking actually pay?	295
Chapter 14. Beyond Plain Text - Tables, Layout, Documents	300
$14.1$ Semi-structured data and why splitting a table corrupts it	300
$14.2$ Table encoding: TAPAS and its positional assumptions	304
$14.3$ Row/column order bias and TableFormer’s fix	309
$14.4$ Document layout: blocks, reading order, hierarchical pre-training	313
$14.5$ Layout priors you should be able to name	318

Part V. Indexing and Vector Search

Unit	Printed page
Part V. Indexing and Vector Search	323
Chapter 15. Approximate Nearest Neighbor Search	324
$15.1$ Why exact search dies: the curse of dimensionality	324
$15.2$ From skip lists to HNSW	328
$15.3$ IVF: coarse quantizers, centroids, Voronoi cells	333
$15.4$ LSH and hash-collision retrieval	337
$15.5$ Graph indexes: RNG, Vamana/DiskANN, Annoy	341
$15.6$ Choosing an index: the recall/latency/memory triangle	345
Chapter 16. Compression and Index Economics	351
$16.1$ Product quantization and the k^m/km argument	351
$16.2$ Residual quantization and IVFADC	355
$16.3$ Scalar and binary quantization	360
$16.4$ FAISS, parallelism, and why block structure matters	364
$16.5$ Index memory arithmetic, end to end	369
Chapter 17. Operating a Vector Store	374
$17.1$ Static index, dynamic corpus: tombstones and re-indexing	374
$17.2$ Sharding, replication, partitioning	378
$17.3$ Choosing a vector database: no winner on all axes	383
$17.4$ Metadata: descriptive, structural, generated	387
$17.5$ Pre-filtering, post-filtering, and the over-filtering risk	392
$17.6$ Self-querying: LLM-extracted filter constraints	396

Part VI. Retrieval and Ranking

Unit	Printed page
Part VI. Retrieval and Ranking	402
Chapter 18. Classical IR You Are Expected to Know	403
$18.1$ Boolean and extended Boolean models	403
$18.2$ Vector space model and TF-IDF’s two failures	408
$18.3$ BM25: saturation and length normalization	412
$18.4$ Probabilistic models and the binary independence assumption	417
$18.5$ Why sparse still wins on the long tail	422
Chapter 19. Learning to Rank	427
$19.1$ The LTR setup and why regression is the wrong framing	427
$19.2$ Pointwise: classification and ordinal regression	431
$19.3$ Pairwise: RankSVM, RankNet, LambdaRank	435
$19.4$ Listwise: SoftRank, LambdaMART, SetRank	440
$19.5$ Neural LTR: representation, interaction, hybrid	444
Chapter 20. Dense Retrieval	449
$20.1$ Bi-encoders, DPR, and the efficiency/expressiveness trade	449
$20.2$ Where dense wins and where BM25 wins	453
$20.3$ Hard negatives, gradient norm, and ANCE	457
$20.4$ Unsupervised dense retrieval: Contriever and MoCo	462
$20.5$ Cross-lingual and low-resource retrieval	466
Chapter 21. Learned Sparse and Multi-Vector Retrieval	472
$21.1$ SPLADE: MLM-predicted term importance and implicit expansion	472
$21.2$ FLOPS regularization: making sparsity a trainable objective	476
$21.3$ ColBERT and late interaction	481
$21.4$ The three ablations that show what makes ColBERT work	485
$21.5$ Making multi-vector affordable: ColBERTv2 and PLAID	489
$21.6$ Hybrid retrieval and Reciprocal Rank Fusion	493
Chapter 22. Reranking	499
$22.1$ Why a second stage exists at all	499
$22.2$ Cross-encoders and choosing rerank depth	503
$22.3$ LLM rerankers: pointwise, pairwise, listwise	507
$22.4$ monoT5 -> RankT5: from classification to real ranking losses	512
$22.5$ Zero-shot listwise prompting and the sliding window	516
$22.6$ Instruction-aware reranking	521
$22.7$ Document-bound vs model-bound	525
Chapter 23. Generative Retrieval	530
$23.1$ Generating document IDs instead of searching	530
$23.2$ Constant-time retrieval vs quality ceiling	534
$23.3$ Incremental indexing: the open wound	538
$23.4$ Recite-then-answer, and GenIR as RAG’s complement	543

Part VII. Query Understanding and Control Flow

Unit	Printed page
Part VII. Query Understanding and Control Flow	548
Chapter 24. Query Reformulation	549
$24.1$ Why queries fail: misspelling, synonymy, polysemy	549
$24.2$ Expansion: lexical, semantic, signal-based	553
$24.3$ Rewriting: Rocchio -> seq2seq -> reinforcement learning	558
$24.4$ Retrieval-specific embeddings: why word2vec is wrong here	562
$24.5$ Conversational rewriting: five operations	566
$24.6$ HyDE: searching with a hypothetical answer	570
Chapter 25. Routing and Adaptive Retrieval	575
$25.1$ When to retrieve: classifier, confidence, and self-reflective families	575
$25.2$ Unified Active Retrieval: four criteria as a decision tree	579
$25.3$ Where to retrieve: source selection, conflict, heterogeneity	583
$25.4$ Query routers: descriptive vs prescriptive	587
$25.5$ Query stratification: four levels of information need	591
$25.6$ Adaptive RAG and silver-label supervision	595
$25.7$ The out-of-distribution weakness every router shares	599
Chapter 26. Iterative, Recursive, and Agentic Retrieval	604
$26.1$ Why one-shot retrieval breaks: three failure modes	604
$26.2$ Iterative: alternating retrieval and generation	608
$26.3$ Recursive: IRCoT and Self-Ask	612
$26.4$ Confidence-triggered: FLARE’s implicit and explicit queries	616
$26.5$ Corrective RAG: evaluate, decompose, fall back to web search	620
$26.6$ Self-RAG: reflection tokens and distilled supervision	624
$26.7$ RL for search: Search-R1, GRPO, and learned stopping	627
$26.8$ Agentic RAG and its costs	632
$26.9$ When to stop: the accuracy/latency/context trade-off	636

Part VIII. Training the RAG System

Unit	Printed page
Part VIII. Training the RAG System	642
Chapter 27. Fine-tuning the Generator	643
$27.1$ SFT teaches behavior, RAG supplies knowledge	643
$27.2$ RAFT: training with distractors on purpose	647
$27.3$ Robustness to retrieval defects: noisy, irrelevant, counterfactual	651
$27.4$ Answer invariance to context as a training objective	655
$27.5$ PEFT: LoRA, QLoRA, and quantization	659
Chapter 28. Training the Retriever	664
$28.1$ Four quadrants: freeze or train x retriever or generator	664
$28.2$ Fusion-in-Decoder and FiD-light	668
$28.3$ Relevance is discrete: Gumbel noise and the differentiability fix	672
$28.4$ REPLUG: black-box LLM, ensembled output distributions	677
$28.5$ LM-supervised retrieval: KL between retriever and LM preference	681
$28.6$ Index refresh during training	686
$28.7$ REALM and retrieval inside pre-training	690
Chapter 29. Bootstrapping Training Data	695
$29.1$ Inverting the relation: what queries does this document answer?	695
$29.2$ doc2query and vocabulary mismatch	699
$29.3$ InPars: contrasting good and bad generated queries	702
$29.4$ Promptagator: retrieval tasks are not one task	706
$29.5$ Consistency filtering	709
$29.6$ Multilingual and conversational extensions	713
$29.7$ Synthetic data: the right question is how, not whether	717
$29.8$ Gold, silver, bronze: label provenance and the teacher ceiling	721

Part IX. Generation and Context Assembly

Unit	Printed page
Part IX. Generation and Context Assembly	727
Chapter 30. Lost in the Middle	728
$30.1$ The U-curve and the serial position effect	728
$30.2$ Worse than closed-book: the finding that changes designs	732
$30.3$ Longer context windows do not help	735
$30.4$ Where the bias comes from: system messages and next-token locality	739
$30.5$ Mitigation by prompting and reordering	743
$30.6$ Ranked-list truncation: retrieve fewer documents	747
$30.7$ Positional encoding and attention calibration	750
$30.8$ Training and architecture fixes: IN2, FILM, DIFF Transformer	753
Chapter 31. Attribution and Citation	758
$31.1$ Attribution as RAG’s core promise	758
$31.2$ The AIS framework: interpretability then attribution	761
$31.3$ Post-hoc citations and why they are nearly worthless	765
$31.4$ Attributable, extrapolatory, contradictory - and the fact-checking mapping	769
$31.5$ How good are automatic attribution judges?	772

Part X. Evaluation

Unit	Printed page
Part X. Evaluation	777
Chapter 32. Evaluating Retrieval	778
$32.1$ Set metrics vs rank metrics	778
$32.2$ MRR, nDCG, recall@k - derived, not quoted	781
$32.3$ Evaluating without ground truth: humans and LLM judges	785
$32.4$ Judge design: exemplars, class balance, macro-F1	788
Chapter 33. Evaluating Generation	793
$33.1$ Faithfulness vs factuality, restated for evaluation	793
$33.2$ FActScore and atomic facts	797
$33.3$ VeriScore: unverifiable claims and the recall correction	802
$33.4$ RAGAS: faithfulness, answer relevance, context relevance	806
$33.5$ Error length dependence and summarization’s open question	810
Chapter 34. Evaluating the System	815
$34.1$ RAG is not end-to-end: attributing failure to a stage	815
$34.2$ Macro, micro, mezzo analysis	818
$34.3$ Ablation design and sanity cases	822
$34.4$ The remove-the-evidence ablation	826
$34.5$ System-level properties: latency, self-consistency, robustness	830
$34.6$ Benchmark hygiene: contamination, synthetic sets, VQA-in-RAG-clothing	834

Part XI. Trust, Credibility, and Adversarial Robustness

Unit	Printed page
Part XI. Trust, Credibility, and Adversarial Robustness	838
Chapter 35. Source Credibility	839
$35.1$ Six aspects across three stages: a trustworthiness taxonomy	839
$35.2$ One bad document among many good ones	842
$35.3$ RA-RAG: reliability from cross-source agreement	845
$35.4$ CAG: training the model to reason about credibility	849
$35.5$ CrAM: down-weighting attention on gullible heads	852
$35.6$ Choosing among the three	856
$35.7$ Temporal awareness and credibility laundering	859
$35.8$ Pluralism: when every credible source shares a blind spot	862
$35.9$ Fair exposure: doubly stochastic rankings and group constraints	866
Chapter 36. Provenance and Adversarial Robustness	870
$36.1$ Attribution is linguistic, authenticity is technical	870
$36.2$ Content credentials and C2PA	873
$36.3$ Missing and forged credentials	877
$36.4$ Indirect prompt injection: the native RAG threat model	881
$36.5$ Data poisoning the datastore	885
$36.6$ Jailbreaking, red teaming, guardrails	889
$36.7$ Adversarial training: discrete, continuous, and hybrid	892
$36.8$ Privacy: what the datastore protects and what it doesn’t	896

Part XII. Scaling, Advanced Variants, and Design Drills

Unit	Printed page
Part XII. Scaling, Advanced Variants, and Design Drills	901
Chapter 37. Latency, Cost, and Systems	902
$37.1$ Where the milliseconds actually go	902
$37.2$ KV-cache reuse for retrieved chunks	906
$37.3$ Reordering a decoder to make caching possible	909
$37.4$ Kernel fusion, static layouts, and GPU-shaped design	913
$37.5$ CPU/GPU pipelining and prefetching	917
$37.6$ Funnel retrieval: coarse-to-fine cascades	921
$37.7$ Recursive summarization for thematic queries	925
$37.8$ Cost per query, end to end	929
Chapter 38. Distributed and Federated RAG	933
$38.1$ Why there is no single vector database	933
$38.2$ Federated search and source selection	936
$38.3$ RAGRoute: learning which sources to query	940
$38.4$ Multi-tenancy, compliance, and data residency	943
$38.5$ RAG’s externalities: the commons and the energy bill	948
Chapter 39. Multimodal RAG	952
$39.1$ Three generations, and the multilingual analogy	952
$39.2$ Encoding: joint vs decoupled-and-aligned	955
$39.3$ CLIP’s three biases	959
$39.4$ Per-modality thresholds	963
$39.5$ Multi-granularity noise correspondence and reranking	966
$39.6$ Fusion: score, attention, unified	970
$39.7$ The time dimension: key frames, hierarchy, causality	974
$39.8$ Generation and non-textual attribution	978
$39.9$ Benchmarks and what they actually measure	982
Chapter 40. Graph RAG	987
$40.1$ When a graph earns its place - and when it doesn’t	987
$40.2$ Constructing the graph: extraction, cost, and quality	991
$40.3$ Query processing: five sub-processes and grounded decomposition	994
$40.4$ Heuristic retrieval: linking, matching, traversal	998
$40.5$ Knowledge graphs as reasoning sources	1002
$40.6$ Learned retrieval: GNN-RAG and query-conditioned message passing	1006
$40.7$ Organizing: pruning, reranking, augmenting, verbalizing	1010
$40.8$ Generation: verbalize, fuse, or stay in graph space	1013
$40.9$ GraphRAG vs LightRAG: communities vs dual-level keys	1017
Chapter 41. End-to-End Design Drills	1022
$41.1$ Enterprise RAG over 100M documents at 500 QPS	1022
$41.2$ A RAG system that must stay fresh	1025
$41.3$ High-stakes RAG: medicine and law	1029
$41.4$ Multi-hop question answering	1033
$41.5$ Multimodal enterprise search	1038
$41.6$ Debugging a RAG system that got worse	1041
$41.7$ Retrofitting credibility onto an existing pipeline	1045

Appendices

Unit	Printed page
Appendix A. Formula Sheet	1050
A.1 BM25	1050
A.2 Reciprocal Rank Fusion	1051
A.3 Mean Reciprocal Rank	1052
A.4 nDCG	1053
A.5 Modularity	1054
A.6 Product Quantization Sizing	1055
A.7 FLOPs and Latency	1056
A.8 Index Memory, End to End	1057
Appendix B. Question Bank	1060
B.1 Core Questions	1060
B.2 Senior Questions	1062
B.3 Staff Questions	1065
B.4 Answers	1068
Appendix C. Design Checklists	1078
C.1 The Retrieval Design Checklist	1078
C.2 The Indexing and Capacity Checklist	1079
C.3 The Evaluation Checklist	1080
C.4 The Credibility and Robustness Checklist	1081
C.5 The Latency and Cost Checklist	1083
Appendix D. The RAG Card and Index Datasheet	1084
D.1 The RAG Card	1084
D.2 The Index Datasheet	1086
Appendix E. Annotated Reading List	1089
E.1 Reading List, Mapped to Chapters	1089
Appendix F. Notation Quick Reference	1101
Appendix G. Glossary	1103
G.1 The Same Idea Under Four Names	1103
G.2 Glossary of Terms	1105

Diagrams

The manifest records zero figures and zero tables for this unit, so there are no source visuals or captions to recreate.

Interview route

Front matter
     |
Part I: frame and design method
     |
Part II -> Part III: generator and context
     |
Part IV -> Part V -> Part VI -> Part VII
representation -> indexing -> ranking -> control flow
     |
Part VIII -> Part IX: training and assembly
     |
Part X -> Part XI: evaluation and trust
     |
Part XII: systems, variants, and design drills
     |
Appendices A through G: rehearsal and reference

Whiteboard pack

Numbered drawing order

Write RAG interview at the top and draw a vertical route beneath it.
Add Part I as the framing station.
Add paired stations for Parts II and III, then Parts IV through VII as the retrieval pipeline.
Add Parts VIII and IX for training and context assembly.
Add Parts X and XI as the evaluation and trust checkpoint.
Add Part XII as the systems and design-drill terminus.
Draw a side box for Appendices A through G and label it formulas, questions, checklists, documentation, readings, notation, glossary.

90-100 word script

The book is organized like a system walkthrough. I start with the interview frame and the generator's limits. Then I move through prompting, representation, indexing, retrieval, and query control because each stage changes what evidence reaches the model. After that, I cover training and context assembly, then evaluate retrieval, generation, and the whole system separately. I finish with credibility, adversarial robustness, scaling, distributed designs, multimodal and graph variants, and end-to-end drills. The appendices give formulas, questions, checklists, documentation templates, readings, notation, and a glossary. This route lets me answer from first principles, then test the design.

Interview traps

1. Why is the contents more useful as a map than as a memorization list?

The outline groups each interview topic with its mechanism and page anchor. Part I frames the answer, Parts II through IX walk through the system, and Parts X and XI test evaluation and trust. Part XII applies the earlier material to advanced variants and design drills.

2. Why not start with agentic, multimodal, or graph retrieval?

The source places those variants after generator behavior, prompting, representation, indexing, ranking, routing, training, context assembly, evaluation, and trust. The order makes the prerequisite design choices visible before Part XII begins on page 901.

3. Where does the outline separate the three evaluation levels?

Part X assigns Chapter 32 to retrieval evaluation on page 778, Chapter 33 to generation evaluation on page 793, and Chapter 34 to system evaluation on page 815. A complete answer should preserve those three levels.

4. Is attribution the whole trust story?

No. Chapter 31 covers attribution and citation on page 758. Part XI separates source credibility on page 839 from provenance and adversarial robustness on page 870, while Section $36.8$ places privacy on page 896.

5. How should an appendix entry connect back to the main outline?

Use an appendix as a reference, then return to the chapter developing its assumptions. The source contents place Appendix A at printed page 1050, Appendix C at 1078 and Appendix D at 1084. Those are preserved source page labels, not page numbers generated by this revised Notion copy.

Key numbers

Number	Source-grounded meaning
10	Portable Document Format (PDF) pages in this contents unit, labeled i through x.
Seven listed front-matter units	The source contents list Preface through About the Author; title and contents themselves are excluded from this seven-unit count.
12	Major parts in the main book.
41	Numbered chapters.
7	Appendices, A through G.
$10^8$ documents	Assumed corpus size in design drill $41.1$ , not a measured deployed scale.
$500\,\mathrm{queries\,s^{-1}}$	Assumed request rate in design drill $41.1$ ; size per-query work, capacity and headroom separately.
1105	Last listed printed page start (G.2); not proof that the book has exactly 1105 total pages.

00-START-HERE · 01-READING-ROUTES · 00_01_title_page · 00_03_preface