Reading routes

00-START-HERE These routes preserve the supplied chapter order as study recommendations, not deployment or interview-readiness certifications. A chapter link opens its Notion page; a decimal label identifies a section to find there. The express route and five role routes follow 00_04_how_to_use_this_book; broader routes follow 98-SOURCE-INDEX. Adapt prerequisites to prior knowledge and fill role-critical gaps. The recommended shared foundation is Part I, Chapters 1 through 3, not the start/index page itself. 00-START-HERE is navigation. Targeted routes can begin later for readers who already know the prerequisites; no route order is a logical guarantee of readiness.

One-week interview path

This deliberately narrow one-week route contains nine selected chapters and two timed scenarios. The last two days are drill time within the week, not additional days. It omits other material that may be critical for a role or deployment.

1. Establish the answer framework with 03_a_repeatable_framework_for_any_rag_design_question and the diagnostic framework with 34_evaluating_the_system.
2. Read the core sequence: 01_what_a_rag_interview_actually_tests → 15_approximate_nearest_neighbor_search → 18_classical_ir_you_are_expected_to_know → 22_reranking → 25_routing_and_adaptive_retrieval → 30_lost_in_the_middle → 32_evaluating_retrieval.
3. Use the last two days for Sections $41.1$ (enterprise sizing) and $41.6$ (regression diagnosis) in 41_end_to_end_design_drills. Treat the mock timing as practice, not a competence cutoff.

Retrieval and ranking quality

This route serves search-relevance and applied-science preparation. It follows the question of why the right document is missing or ranked too low.

Vector search and platform infrastructure

This route serves index owners who must defend capacity, recall targets, latency budgets, and operational choices. 03_a_repeatable_framework_for_any_rag_design_question → 15_approximate_nearest_neighbor_search → 16_compression_and_index_economics → 17_operating_a_vector_store → 37_latency_cost_and_systems → 38_distributed_and_federated_rag, then Sections $41.1$ (enterprise sizing) and $41.2$ (freshness) in 41_end_to_end_design_drills. Begin Chapter 3 with Section $3.4$ on sizing. Begin with workload and capacity assumptions, then index choice, scoped compression budgets and operations. Continue through latency and federation before the two drills. Corpus size and request rate jointly affect resources; a payload subtotal alone is not a node count.

Shipping a RAG product

This route serves applied-AI and product-engineering preparation. It follows the complete product loop and failures that affect users. 01_what_a_rag_interview_actually_tests → 03_a_repeatable_framework_for_any_rag_design_question → 04_what_the_llm_brings_and_what_it_breaks → 09_prompting_for_retrieval_augmented_systems → 11_abstention_and_calibration → 13_chunking_and_granularity → 25_routing_and_adaptive_retrieval → 30_lost_in_the_middle → 31_attribution_and_citation → 33_evaluating_generation → 34_evaluating_the_system, then Section $41.6$ in 41_end_to_end_design_drills. Within 09_prompting_for_retrieval_augmented_systems, Section $9.4$ examines conditions under which more retrieved material can hurt generated answers. The final Section $41.6$ drill tests regression hypotheses; it does not guarantee a unique responsible stage.[1][2]

High-stakes, trust, and safety

This route supports study for regulated settings and unreliable/hostile sources. Use Part I prerequisites as needed. Reading a route does not establish clinical, legal, privacy or security compliance. 04_what_the_llm_brings_and_what_it_breaks → 05_data_privacy_and_the_legal_surface → 11_abstention_and_calibration → 31_attribution_and_citation → 35_source_credibility → 36_provenance_and_adversarial_robustness, then Sections $41.3$ (high-stakes use) and $41.7$ (credibility retrofit) in 41_end_to_end_design_drills. Connect generator failures and addressable data boundaries to abstention, attribution, credibility and adversarial tests. Datastore removal handles do not establish legal immunity, and citations or provenance alone do not prove truth. The two drills make these distinctions under explicit assumptions.[3][4][5]

Training the system and advanced variants

This route serves as research-oriented preparation on how to change the components of a RAG system. It assumes the Part I foundation. 06_scaling_laws_and_the_economics_of_retrieval_vs_parameters → 07_in_context_learning_the_mechanism_rag_rides_on → 08_reading_the_machine_circuits_induction_heads_and_attribution → 27_fine_tuning_the_generator → 28_training_the_retriever → 29_bootstrapping_training_data → 23_generative_retrieval → 26_iterative_recursive_and_agentic_retrieval → 39_multimodal_rag → 40_graph_rag. You begin with retrieval economics and in-context learning, then study interpretability and training. The route continues through generative and iterative retrieval before multimodal and graph variants.

Production system path

This production-study route moves from framing to operational review. Add Chapters 39 and 40 when modality, graph relationships, provenance or global queries change the problem. It is a reading map, not a complete release checklist or substitute for domain-specific safety and access requirements.

1. You establish system and ingestion decisions with 03_a_repeatable_framework_for_any_rag_design_question → 12_text_representation → 13_chunking_and_granularity → 14_beyond_plain_text_tables_layout_documents.
2. You size and operate retrieval with 15_approximate_nearest_neighbor_search → 16_compression_and_index_economics → 17_operating_a_vector_store → 21_learned_sparse_and_multi_vector_retrieval → 22_reranking.
3. You control queries and context with 24_query_reformulation → 25_routing_and_adaptive_retrieval → 26_iterative_recursive_and_agentic_retrieval → 30_lost_in_the_middle.
4. You validate behavior and operations with 31_attribution_and_citation → 32_evaluating_retrieval → 33_evaluating_generation → 34_evaluating_the_system → 35_source_credibility → 36_provenance_and_adversarial_robustness → 37_latency_cost_and_systems → 38_distributed_and_federated_rag → A4_the_rag_card_and_index_datasheet.
5. You add 39_multimodal_rag and 40_graph_rag when multimodal or graph requirements apply.
6. You finish the design review with 41_end_to_end_design_drills and A3_design_checklists.

Broad interview loop

This is the longer fast-interview loop from the repository index. It covers more of the retrieval, evaluation, and systems stack than the separate one-week path.

1. You calibrate interview expectations with 00_05_for_interview_candidates → 01_what_a_rag_interview_actually_tests → 03_a_repeatable_framework_for_any_rag_design_question.
2. You review vector search and classical retrieval with 15_approximate_nearest_neighbor_search → 16_compression_and_index_economics → 17_operating_a_vector_store → 18_classical_ir_you_are_expected_to_know.
3. You study hybrid retrieval and query control with 21_learned_sparse_and_multi_vector_retrieval → 22_reranking → 24_query_reformulation → 25_routing_and_adaptive_retrieval → 26_iterative_recursive_and_agentic_retrieval.
4. You review context use and evaluation with 30_lost_in_the_middle → 31_attribution_and_citation → 32_evaluating_retrieval → 33_evaluating_generation → 34_evaluating_the_system.
5. You cover credibility, provenance, systems, and federation with 35_source_credibility → 36_provenance_and_adversarial_robustness → 37_latency_cost_and_systems → 38_distributed_and_federated_rag.
6. You finish with 41_end_to_end_design_drills, then rehearse with A1_formula_sheet, A2_question_bank, A3_design_checklists, and A6_notation_quick_reference.

Complete technical path

The complete content sequence is available from 00-START-HERE. The groups below are navigation shortcuts, not instructions to reread the start page in place of the chapters.

1. Read the nine front-matter units, especially 00_04_how_to_use_this_book and 00_07_notation_and_symbols.
2. Read Chapters 1 through 14, from 01_what_a_rag_interview_actually_tests to 14_beyond_plain_text_tables_layout_documents, for framing, model behavior, prompting and representation.
3. Read Chapters 15 through 29, from 15_approximate_nearest_neighbor_search to 29_bootstrapping_training_data, for indexing, ranking, control and training.
4. Read Chapters 30 through 41, from 30_lost_in_the_middle to 41_end_to_end_design_drills, for context, evaluation, trust, systems and drills.
5. Use Appendices A1 through A7, listed in the appendix map below, for reference and practice. 00-START-HERE links every content unit.

Evaluation and trust path

The source index also provides this focused evaluation/trust study route. Finish a real handoff with measured artifacts and domain-specific review; completing the listed reading is not by itself release clearance.

1. You establish data boundaries, prompt sensitivity, and abstention with 05_data_privacy_and_the_legal_surface → 10_prompt_sensitivity → 11_abstention_and_calibration.
2. You study training limits with 27_fine_tuning_the_generator → 28_training_the_retriever → 29_bootstrapping_training_data.
3. You connect attribution to evaluation and trust with 31_attribution_and_citation → 32_evaluating_retrieval → 33_evaluating_generation → 34_evaluating_the_system → 35_source_credibility → 36_provenance_and_adversarial_robustness.
4. You examine operating costs and distributed governance with 37_latency_cost_and_systems → 38_distributed_and_federated_rag.
5. You complete A4_the_rag_card_and_index_datasheet for the production handoff.

Design drills

Chapter 41 contains seven scenarios. Each link below opens the same chapter; the explicit section label identifies the scenario to find there. The workloads and numerical budgets are assumptions for the drills, not deployment benchmarks.

Study routine

Recurring teaching blocks support the six-step study routine below, but layouts and source coverage vary by unit. This six-step routine is distinct from the seven-element teaching template described in the usage guide. Re-derive assumptions and units, and allow additional time for equations, code, exercises and empirical checks.

Core, Senior and Staff are the source’s practice tags for mechanisms, constrained derivation and judgment under changed conditions; they are not independently calibrated hiring levels. Practice with A2_question_bank and review with A3_design_checklists. A useful study checkpoint is an explanation with a choice, definitions, units, cost, failure mode and reversal condition. Unaided recall is one practice tool, not proof of correctness; verify with references, calculations and tests when making engineering decisions.

Appendix map

Source references to Appendix A through Appendix G correspond to the following repository files. For example, the source Formula Sheet is Appendix A, while its local file identifier is A1.

00-START-HERE