Appendix E: Annotated Reading List

00-START-HERE · 01-READING-ROUTES · A4_the_rag_card_and_index_datasheet · A6_notation_quick_reference Use this curated, chapter-mapped reading list to connect each Retrieval-Augmented Generation (RAG) paper with the exact result a senior interview candidate should be able to reproduce.

TL;DR

• The list contains $161$ entries in $14$ clusters and $52$ core markers, covering $160$ distinct titles; Cumulated Gain-Based Evaluation appears in two clusters. The source estimates roughly $500$ cited works overall, a broader count not independently re-enumerated here.
• Learn the result together with its model, task, baseline, assumptions and uncertainty. An annotation is not a universal theorem.
• Preserve the supplied paper titles, author strings and priority markers; chapter links guide interpretation and do not replace the primary paper.
• Negative results are conditional evidence about failure modes, not guarantees that a method always fails.
• Read the ten-paper starter route for orientation, then follow the clusters relevant to the actual role or system.

The story

Imagine fourteen shelves containing $161$ reading cards for $160$ distinct titles. One evaluation paper appears twice because it serves two chapter clusters. The source’s roughly five hundred cited works describe the wider bibliography, not a newly verified unique-work count in this appendix. Fifty-two cards carry a [core] seal. The seal does not mean that the other cards are unimportant. It means a senior visitor should explain the sealed card without asking the curator for a hint. The curator refuses title collecting. If a visitor says “Lost in the Middle,” the curator asks what baseline it used and what the comparison showed. A title is only a shelf label. The annotation is the knowledge that must leave the room with the visitor. Several shelves hold uncomfortable findings. Retrieval can lower accuracy. Semantic chunking gains may not transfer. Automatic citation judges can disagree with people. Self-critique without outside feedback may not improve an answer. The curator keeps these cards near the front because good interviews probe limits, not just success stories. Some cards seem to belong in neighboring libraries. Distance concentration, tail latency, and instruction hierarchies are not retrieval papers. The curator keeps them because those constraints still govern the RAG reading room. A visitor with limited time follows the ten-card route. It introduces the book’s design decisions, selected negative results, index arithmetic and a threat model; it does not certify complete preparation for every senior interview.

Decoder table

Core mechanics

Use each annotation as a question to investigate in the original paper: what was measured, against which baseline, under which assumptions, and at what cost? Titles and supplied author strings are preserved here; per-entry years and venues are not invented. The chapter citations are interpretation cross-references, not claims that every listed paper was experimentally reproduced during this review. The actual list has $161$ entries, $160$ distinct titles, $14$ clusters and $52$ core markers. The repeated title is Cumulated Gain-Based Evaluation of IR Techniques. The source’s roughly $500$ cited works is a separate estimate of the wider bibliography. For mathematical reading, distinguish exact identities from estimates: $C\approx6ND$ is a training-compute proxy, LoRA counts trainable factors rather than all FLOPs, $4d+8M$ is only a scoped HNSW subtotal, and PQ codebook scalars equal $(km)(d/m)=kd$ for equal subspaces. Read the conditions in the annotations before applying any number. The source gives two usage notes. The annotations emphasize negative results, including retrieval lowering accuracy, semantic chunking gains not transferring, weak citation quality, and unreliable self-correction without external feedback. The list also includes governing constraints from non-retrieval papers, including distance concentration, tail latency, and instruction hierarchies.

1. What RAG is, and when it fails

What. This cluster supports chapter 1 and chapter 3. Why. It defines the pattern, compares parametric and non-parametric knowledge, and includes the central negative result. Failure without it. A candidate can name RAG without explaining its marginalizations, its datastore rationale, or when retrieval hurts. Cost and complexity. The cluster contains 7 entries, including 2 marked [core].

1. Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks [core]. Authors: Lewis, Perez, Piktus et al. Claim to know: The paper named the pattern. Know the parametric / non-parametric split. RAG-Sequence and RAG-Token are different marginalizations, not implementation details.[1]
2. When Not to Trust Language Models: Investigating Effectiveness of Parametric and Non-Parametric Memories [core]. Authors: Mallen, Asai, Zhong et al. Claim to know: In the reported models/tasks, retrieval can change an otherwise-correct closed-book answer into a wrong one; the roughly $10\%$ figure is conditional on that evaluation, not a universal harm rate. Study the popularity-stratified baselines and Wikipedia-page-view proxy.[1]
3. Large Language Models Struggle to Learn Long-Tail Knowledge. Authors: Kandpal, Deng, Roberts et al. Claim to know: The study relates factual accuracy to entity co-occurrence in the studied training data. This is empirical evidence motivating external knowledge, not a universal quantitative law for every model.[1]
4. Nonparametric Masked Language Modeling. Authors: Min, Shi, Lewis et al. Claim to know: Retrieval is a modeling choice rather than a bolt-on. This is the cleanest statement of what non-parametric capacity buys.[1]
5. Toolformer: Language Models Can Teach Themselves to Use Tools. Authors: Schick, Dwivedi-Yu, Dessi et al. Claim to know: It establishes retrieval as one instance of tool use. It is useful for the framing question “why is search special?”[1]
6. ReAct: Synergizing Reasoning and Acting in Language Models. Authors: Yao, Zhao, Yu et al. Claim to know: It supplies the interleaved reason-act loop that later agentic retrieval builds on.[1]
7. Retrieval-Augmented Generation for Large Language Models: A Survey. Authors: Gao, Xiong, Gao et al. Claim to know: Use the naive/advanced/modular taxonomy to describe design patterns, not to certify maturity, quality or an inevitable progression.[1]

2. The alternatives: editing, continual learning, long context

What. This cluster supports chapter 2. Why. It compares retrieval with editing, continual learning, and generative retrieval. Failure without it. A candidate can recommend changing weights without knowing localization limits, ripple failures, forgetting, or indexing wounds. Cost and complexity. The cluster contains 9 entries, including 1 marked [core].

1. Locating and Editing Factual Associations in GPT [core]. Authors: Meng, Bau, Andonian et al. Claim to know: ROME uses causal tracing and rank-one edits to study factual associations in specified models. A causal intervention at middle-layer MLPs does not imply that each fact occupies an isolated, uniquely owned weight location.[2]
2. Mass-Editing Memory in a Transformer. Authors: Meng, Sharma, Andonian et al. Claim to know: MEMIT is the scaling answer to ROME. Know where batch editing starts to degrade unrelated knowledge.[2]
3. Does Localization Inform Editing? Surprising Differences in Causality-Based Localization. Authors: Hase, Bansal, Kim et al. Claim to know: Compare causal-localization scores with edit efficacy under the paper’s methods and models; localization need not reliably predict the best edit site. This is not a theorem that localization never informs any editor.[2]
4. Evaluating the Ripple Effects of Knowledge Editing in Language Models. Authors: Cohen, Biran, Yoran et al. Claim to know: Edits can fail to propagate consistently to related or entailed facts. Inspect the ripple evaluation rather than assuming all edits always fail or always preserve consistency.[2]
5. Overcoming Catastrophic Forgetting in Neural Networks. Authors: Kirkpatrick, Pascanu, Rabinowitz et al. Claim to know: EWC frames forgetting as a constraint on parameter drift.[2]
6. An Empirical Study of Catastrophic Forgetting in Large Language Models During Continual Fine-tuning. Authors: Luo, Yang, Meng et al. Claim to know: Know how much is actually lost and at what scale.[2]
7. Transformer Memory as a Differentiable Search Index. Authors: Tay, Tran, Dehghani et al. Claim to know: DSI generates identifiers with a learned model instead of the usual explicit similarity-search path. Identifier validity, evidence access and corpus updates still incur work; update difficulty is not proof that incremental methods cannot exist.[2]
8. DSI++: Updating Transformer Memory with New Documents. Authors: Mehta, Gupta, Tay et al. Claim to know: Study the proposed new-document update method, its costs and retained-document performance. Its existence rules out a blanket claim that generative retrieval admits no incremental updating.[2]
9. Rethinking Search: Making Domain Experts out of Dilettantes. Authors: Metzler, Tay, Bahri et al. Claim to know: This is the position paper behind generative retrieval. It is useful for the “are GenIR and RAG contradictory?” exchange.[2]

3. What the generator brings, and what it breaks

What. This cluster supports chapter 4. Why. It covers memorization, extraction, hallucination, faithfulness, irrelevant context, and confidence. Failure without it. A candidate can treat generation as a neutral final stage and miss its retained data, calibration limits, and context failures. Cost and complexity. The cluster contains 9 entries, including 3 marked [core].

1. Quantifying Memorization Across Neural Language Models [core]. Authors: Carlini, Ippolito, Jagielski et al. Claim to know: The study reports increased extractable memorization with model scale, duplication and supplied context in its tested regime. These are empirical trends, not three unrestricted probability laws.[3]
2. Scalable Extraction of Training Data from (Production) Language Models. Authors: Nasr, Carlini, Hayase et al. Claim to know: Know the divergence attack mechanism, not merely that it exists.[3]
3. Deduplicating Training Data Makes Language Models Better. Authors: Lee, Ippolito, Nystrom et al. Claim to know: The reported experiments reduce extractable memorization and can improve language-model quality through deduplication. Transfer and magnitude must be checked for the actual corpus and recipe.[3]
4. Survey of Hallucination in Natural Language Generation. Authors: Ji, Lee, Frieske et al. Claim to know: Know the intrinsic-versus-extrinsic taxonomy used by the book and where each type originates.[3]
5. On Faithfulness and Factuality in Abstractive Summarization [core]. Authors: Maynez, Narayan, Bohnet et al. Claim to know: This distinction supports the evaluation half of the book. An answer can be faithful and false.[3]
6. Large Language Models Can Be Easily Distracted by Irrelevant Context [core]. Authors: Shi, Chen, Misra et al. Claim to know: Adding irrelevant passages can degrade accuracy in the tested setup. More retrieved text is not guaranteed to help; the direction depends on content, model and task.[3]
7. Impact of Co-occurrence on Factual Knowledge of Large Language Models. Authors: Kang and Choi. Claim to know: Models can substitute a frequently co-occurring entity for the correct one.[3]
8. SelfCheckGPT: Zero-Resource Black-Box Hallucination Detection. Authors: Manakul, Liusie, Gales. Claim to know: Sampling-based agreement is a fallible hallucination signal, not a truth certificate. Account for sample cost and shared systematic errors.[3]
9. Language Models (Mostly) Know What They Know. Authors: Kadavath, Conerly, Askell et al. Claim to know: Self-evaluation is partially calibrated. This is both the foundation and the limit of confidence-based gating.[3]

4. Privacy, law, and the argument for a datastore

What. This cluster supports chapter 5. Why. It ties corpus composition, legal risk, toxicity, and training-data extraction to deployment design. Failure without it. A candidate can make a datastore recommendation without its legal architecture or empirical corpus rationale. Cost and complexity. The cluster contains 3 entries, including 1 marked [core].

1. SILO Language Models: Isolating Legal Risk in a Nonparametric Datastore [core]. Authors: Min, Gururangan, Wallace et al. Claim to know: SILO trains on selected lower-risk tiers and keeps revocable material keyed in a datastore. This makes later serving exclusion addressable; it does not itself establish the lawfulness of acquisition, indexing or every later use.[4]
2. RealToxicityPrompts: Evaluating Neural Toxic Degeneration in Language Models. Authors: Gehman, Gururangan, Sap et al. Claim to know: Know what is measurably present in a web-scraped corpus.[4]
3. Extracting Training Data from Large Language Models. Authors: Carlini, Tramer, Wallace et al. Claim to know: The extraction result made corpus composition a deployment question rather than a research question.[4]

5. Scaling, economics, and in-context learning

What. This cluster supports chapter 6 and chapter 7. Why. It supplies scaling arithmetic, compute-optimal practice, in-context behavior, calibration, model configuration, and adaptation cost. Failure without it. A candidate can quote scaling or fine-tuning claims without using their arithmetic or knowing what mechanism changes. Cost and complexity. The cluster contains 7 entries, including 4 marked [core].

1. Scaling Laws for Neural Language Models [core]. Authors: Kaplan, McCandlish, Henighan et al. Claim to know: Use $C\approx6ND$ as a scoped dense-model training-FLOP proxy: $N$ is parameter count and $D$ cumulative processed training tokens. Study the original scaling regime separately from later compute-optimal results; attention, utilization and system overhead are not represented by this proxy.[5]
2. Training Compute-Optimal Large Language Models [core]. Authors: Hoffmann, Borgeaud, Mensch et al. Claim to know: Chinchilla studies a compute-optimal model/data allocation under its training setup. The often-quoted roughly twenty tokens per parameter is not a universal constant; lifetime inference demand can change the economic optimum.[5]
3. Language Models are Few-Shot Learners [core]. Authors: Brown, Mann, Ryder et al. Claim to know: This is the mechanism RAG rides on. In-context learning is conditioning, not a weight update.[5]
4. Rethinking the Role of Demonstrations: What Makes In-Context Learning Work?. Authors: Min, Lyu, Holtzman et al. Claim to know: In the paper’s tasks, demonstration distribution and label-space cues can matter strongly even when labels are corrupted. Do not generalize this into “label correctness never matters,” especially for task learning or safety-critical examples.[5]
5. Calibrate Before Use: Improving Few-Shot Performance of Language Models. Authors: Zhao, Wallace, Feng et al. Claim to know: Content-free calibration, and the ordering sensitivity that makes it necessary.[5]
6. The Llama 3 Herd of Models. Authors: Grattafiori, Dubey et al. Claim to know: This is the source of the book’s default configuration numbers, including layer count, KV head count, head dimension, and the token budget used in memory arithmetic.[5]
7. LoRA: Low-Rank Adaptation of Large Language Models [core]. Authors: Hu, Shen, Wallis et al. Claim to know: For one adapted matrix, two rank-$r$ factors add $r(d_{\mathrm{in}}+d_{\mathrm{out}})$ trainable parameters, excluding biases and other adaptations. Fewer trainable parameters reduce optimizer/state costs but do not remove the frozen base-model forward and backward-path computation.[5]

6. Reading the machine

What. This cluster supports chapter 8. Why. It covers transformer circuits, induction, parameter storage, feature attribution, and the limits of attention as evidence. Failure without it. A candidate can use mechanistic language without knowing which evidence supports information flow, storage, or attribution. Cost and complexity. The cluster contains 5 entries, including 1 marked [core].

1. A Mathematical Framework for Transformer Circuits [core]. Authors: Elhage, Nanda, Olsson et al. Claim to know: The residual stream is an information bus. Know the $QK$/$OV$ factorization, which circuit decides where to look, and which decides what to move.[6]
2. In-context Learning and Induction Heads. Authors: Olsson, Elhage, Nanda et al. Claim to know: Induction heads provide a copying mechanism and evidence about some in-context behavior in the studied transformers. They are not a complete or universal causal explanation of retrieval-augmented generation.[6]
3. Transformer Feed-Forward Layers Are Key-Value Memories. Authors: Geva, Schuster, Berant et al. Claim to know: The key-value-memory interpretation of feed-forward layers connects to model-editing experiments. It does not identify an exclusive location for all facts or an independent fact-key delete.[6]
4. Axiomatic Attribution for Deep Networks. Authors: Sundararajan, Taly, Yan. Claim to know: Integrated gradients are the reference point for what feature attribution can and cannot claim about a generated answer.[6]
5. Attention is not Explanation. Authors: Jain and Wallace. Claim to know: The paper challenges treating attention weights alone as faithful explanations in its evaluated settings. Attention visualizations require causal/behavioral validation; this is not proof that every attention-based analysis is useless.[6]

7. Prompting, sensitivity, and abstention

What. This cluster supports chapter 9 through chapter 11. Why. It covers reasoning scaffolds, prompt search, calibration, uncertainty, self-correction, and recursive-data failure. Failure without it. A candidate can recommend more reasoning or self-critique without pricing samples, measuring sensitivity, or checking external feedback. Cost and complexity. The cluster contains 11 entries, including 3 marked [core].

1. Chain-of-Thought Prompting Elicits Reasoning in Large Language Models [core]. Authors: Wei, Wang, Schuurmans et al. Claim to know: This is the scaffold everything else in the cluster modifies.[7]
2. Self-Consistency Improves Chain of Thought Reasoning in Language Models. Authors: Wang, Wei, Schuurmans et al. Claim to know: Sample reasoning paths and aggregate answers. Gains depend on task, error dependence and voting rules; separate total generation work from latency with shared prefixes or parallel sampling.[7]
3. Tree of Thoughts: Deliberate Problem Solving with Large Language Models. Authors: Yao, Yu, Zhao et al. Claim to know: Search over reasoning states, and where the added scaffolding stops paying.[7]
4. POSIX: A Prompt Sensitivity Index for Large Language Models. Authors: Chatterjee et al. Claim to know: A measurable definition of prompt sensitivity makes it an engineering variable rather than folklore.[7]
5. Large Language Models Are Human-Level Prompt Engineers. Authors: Zhou, Muresanu, Han et al. Claim to know: Automated prompt search, and the baseline any hand-tuned prompt should beat.[7]
6. On Calibration of Modern Neural Networks [core]. Authors: Guo, Pleiss, Sun, Weinberger. Claim to know: Expected calibration error, and the observation that larger models are not automatically better calibrated.[7]
7. Semantic Uncertainty: Linguistic Invariances for Uncertainty Estimation. Authors: Kuhn, Gal, Farquhar. Claim to know: Group outputs by estimated meaning before measuring entropy. This can avoid treating paraphrases as disagreement, but semantic clustering can err and low entropy can reflect a shared false answer.[7]
8. Detecting Hallucinations in Large Language Models Using Semantic Entropy. Authors: Farquhar, Kossen, Kuhn et al. Claim to know: Study semantic entropy as an uncertainty signal and validate its failure prediction on the deployment task. It is not a general hallucination-proofing guarantee.[7]
9. Just Ask for Calibration. Authors: Tian, Mitchell, Zhou et al. Claim to know: Verbalized confidence from a black-box model, and how far it can be trusted.[7]
10. Large Language Models Cannot Self-Correct Reasoning Yet [core]. Authors: Huang, Chen, Mishra et al. Claim to know: Self-critique without external feedback need not reliably improve reasoning in the tested settings. Compare feedback access, stopping, model and task when reading Self-Refine or Reflexion; differing conditions need not produce a contradiction.[7]
11. AI Models Collapse When Trained on Recursively Generated Data. Authors: Shumailov, Shumaylov, Zhao et al. Claim to know: The paper studies recursive training-data contamination by model-generated samples. It does not prove that an inference-time self-reflection loop automatically causes the same training-distribution collapse.[7]

8. Representation, chunking, and document structure

What. This cluster supports chapter 12 through chapter 14. Why. It connects embedding objectives, evaluation, granularity, chunking, tables, and page layout. Failure without it. A candidate can select embeddings or chunks by leaderboard and label without checking task dependence, anisotropy, transfer, or structure. Cost and complexity. The cluster contains 13 entries, including 3 marked [core].

1. Efficient Estimation of Word Representations in Vector Space. Authors: Mikolov, Chen, Corrado, Dean. Claim to know: Study word2vec’s distributional objectives and static representations. Their objective differs from query-document relevance training; that does not make every use of static embeddings invalid.[8]
2. Sentence-BERT: Sentence Embeddings using Siamese BERT-Networks [core]. Authors: Reimers and Gurevych. Claim to know: Independent sentence encoding enables reuse of document vectors, whereas a general joint cross-encoder score cannot be reproduced by one query-independent document vector. Compare the lost joint interaction and measured efficiency; some pair-level work can still be cached.[8]
3. SimCSE: Simple Contrastive Learning of Sentence Embeddings. Authors: Gao, Yao, Chen. Claim to know: Dropout as the minimal augmentation, and the alignment/uniformity framing.[8]
4. MTEB: Massive Text Embedding Benchmark. Authors: Muennighoff, Tazi, Magne, Reimers. Claim to know: This is the benchmark you will be asked about when someone proposes an embedding model. Leaderboard position is task-dependent.[8]
5. How Contextual are Contextualized Word Representations?. Authors: Ethayarajh. Claim to know: Study anisotropy and similarity baselines in the evaluated representations. The reliability of raw cosine depends on the embedding and task, not merely on the existence of anisotropy.[8]
6. Dense X Retrieval: What Retrieval Granularity Should We Use? [core]. Authors: Chen, Zhang, Zhang et al. Claim to know: Propositions as the retrieval unit. This is the clearest statement that granularity is a first-class design decision.[8]
7. Is Semantic Chunking Worth the Computational Cost? [core]. Authors: Qu, Tu, Bao. Claim to know: Audit whether benchmark gains transfer from stitched or otherwise constructed inputs to naturally occurring documents, and compare against matched fixed-size baselines. Do not infer that every semantic chunker is ineffective or that all reported gains share one cause.[8]
8. RAPTOR: Recursive Abstractive Processing for Tree-Organized Retrieval. Authors: Sarthi, Abdullah, Tuli et al. Claim to know: Recursive summaries offer additional granularity and can help global/thematic questions. Measure them against strong flat retrieval; a query label alone does not prove that flat chunks cannot answer it.[8]
9. Mix-of-Granularity: Optimize the Chunking Granularity for RAG. Authors: Zhong, Liu, Wang et al. Claim to know: Route over chunk sizes rather than choosing one.[8]
10. A Critique and Improvement of an Evaluation Metric for Text Segmentation. Authors: Pevzner and Hearst. Claim to know: WindowDiff, and why $P_k$ under-penalizes certain boundary errors.[8]
11. TaPas: Weakly Supervised Table Parsing via Pre-training. Authors: Herzig, Nowak, Muller et al. Claim to know: Row and column position embeddings, and the assumptions they bake in.[8]
12. TableFormer: Robust Transformer Modeling for Table-Text Encoding. Authors: Yang, Gupta, Upadhyay et al. Claim to know: Study the table-structure encoding and the permutations it is designed to handle. Order invariance is architecture- and task-specific; ordered rows and columns can carry real semantics.[8]
13. LayoutLM: Pre-training of Text and Layout for Document Image Understanding. Authors: Xu, Li, Cui et al. Claim to know: Layout as a first-class input rather than a preprocessing step.[8]

9. Vector indexes and compression

What. This cluster supports chapter 15 through chapter 17. Why. It supplies the graph, compression, storage, filtering, benchmarking, dimensionality, and tail-latency foundations. Failure without it. A candidate can quote one latency number without deriving the index, checking recall, or understanding fan-out and high-dimensional limits. Cost and complexity. The cluster contains 15 entries, including 5 marked [core].

1. Skip Lists: A Probabilistic Alternative to Balanced Trees [core]. Authors: Pugh. Claim to know: Pugh derives probabilistic skip-list search bounds, including a form $L(n)/p+1/(1-p)$ under the stated promotion convention. HNSW borrows a hierarchy analogy, not the ordered skip list’s exact comparison bound or every search guarantee.[9]
2. Efficient and Robust Approximate Nearest Neighbor Search Using Hierarchical Navigable Small World Graphs [core]. Authors: Malkov and Yashunin. Claim to know: Know the common $m_L=1/\ln M$ convention for $M\gt 1$, base capacity $M_0=2M$ and neighbor-diversity heuristic. The $4d+8M$ bytes rule includes fp32 vector payload and allocated base-layer links with four-byte IDs only; add upper layers and all other metadata/runtime terms.[9]
3. Approximate Nearest Neighbor Algorithm Based on Navigable Small World Graphs. Authors: Malkov, Ponomarenko, Logvinov, Krylov. Claim to know: Flat NSW, and why the hierarchy was needed.[9]
4. Product Quantization for Nearest Neighbor Search [core]. Authors: Jegou, Douze, Schmid. Claim to know: With $m$ equal subspaces in dimension $d$ and $k$ centroids per subspace, at most $k^m$ reconstructions use $km$ subspace centroids containing $kd$ scalar values. Thus centroid bytes are independent of $m$ only at fixed $k,d$ and value precision; code length still grows with $m$.[9]
5. Optimized Product Quantization for Approximate Nearest Neighbor Search. Authors: Ge, He, Ke, Sun. Claim to know: Learning the rotation before splitting, and when it is worth the extra step.[9]
6. Billion-Scale Similarity Search with GPUs [core]. Authors: Johnson, Douze, Jegou. Claim to know: FAISS. This is the source against which the book sanity-checks its memory rules of thumb.[9]
7. DiskANN: Fast Accurate Billion-point Nearest Neighbor Search on a Single Node. Authors: Subramanya, Devvrit, Kadekodi et al. Claim to know: Vamana, and the SSD-resident regime where the graph no longer has to fit in RAM.[9]
8. FreshDiskANN: A Fast and Accurate Graph-Based ANN Index for Streaming Similarity Search. Authors: Singh, Subramanya, Krishnaswamy, Simhadri. Claim to know: Deletion and insertion in a graph index without a full rebuild. Read it alongside the tombstone material.[9]
9. Similarity Estimation Techniques from Rounding Algorithms. Authors: Charikar. Claim to know: SimHash, and the LSH collision-probability argument.[9]
10. Approximate Nearest Neighbors: Towards Removing the Curse of Dimensionality. Authors: Indyk and Motwani. Claim to know: Study the near/far collision guarantees and the parameter, approximation and output assumptions in LSH query bounds. High dimension alone does not establish a universal performance ranking against every graph index.[9]
11. When Is “Nearest Neighbor” Meaningful? [core]. Authors: Beyer, Goldstein, Ramakrishnan, Shaft. Claim to know: Distance concentration follows under specified distributional conditions and can weaken distance contrast. It neither proves exact search always becomes infeasible nor makes ANN mathematically necessary for every workload.[9]
12. ANN-Benchmarks: A Benchmarking Tool for Approximate Nearest Neighbor Algorithms. Authors: Aumuller, Bernhardsson, Faithfull. Claim to know: The recall-versus-QPS curve as the reporting convention. Ask for this plot whenever a vendor quotes one latency number.[9]
13. Matryoshka Representation Learning. Authors: Kusupati, Bhatt, Rege et al. Claim to know: Nested representations are trained to make selected prefixes useful. Truncation quality is empirical, and arbitrary embeddings do not acquire this property for free.[9]
14. ACORN: Performant and Predicate-Agnostic Search Over Vector Embeddings. Authors: Patel, Kraft, Guestrin, Zaharia. Claim to know: Filtered search without a separate index per predicate, and the over-filtering failure it addresses.[9]
15. The Tail at Scale. Authors: Dean and Barroso. Claim to know: Fan-out can amplify request-level tail latency through the slowest required responses. Dependence, load, retries and quorum policy matter; measure the actual $p_{99}$ rather than assuming a universal shard formula.[9]

10. Ranking: lexical, dense, sparse, multi-vector

What. This cluster supports chapter 18 through chapter 23. Why. It covers lexical ranking, dense and sparse training, late interaction, reranking, learning to rank, evaluation, datasets, and fusion. Failure without it. A candidate can default to dense-only retrieval without its out-of-domain counterexample, storage cost, benchmark bias, or ranking metrics. Cost and complexity. The cluster contains 17 entries, including 8 marked [core].

1. The Probabilistic Relevance Framework: BM25 and Beyond [core]. Authors: Robertson and Zaragoza. Claim to know: This is the reference for BM25. Know where $k_1$ and $b$ enter and why the length discount sits inside the saturation denominator.[10]
2. Okapi at TREC-3. Authors: Robertson, Walker, Jones et al. Claim to know: Read the original evaluation and parameterization. Popular defaults are implementation and collection choices, not universally optimal constants.[10]
3. Pivoted Document Length Normalization. Authors: Singhal, Buckley, Mitra. Claim to know: Why length normalization was needed at all, shown empirically.[10]
4. A Statistical Interpretation of Term Specificity. Authors: Sparck Jones. Claim to know: IDF’s origin. This is short and still the clearest statement of the intuition.[10]
5. Dense Passage Retrieval for Open-Domain Question Answering [core]. Authors: Karpukhin, Oguz, Min et al. Claim to know: Know the in-batch-negative setup, positive-label assumptions and independently encoded query/document efficiency argument. Check the evaluation protocol rather than relying on a citation-count superlative.[10]
6. Approximate Nearest Neighbor Negative Contrastive Learning [core]. Authors: Xiong, Xiong, Li et al. Claim to know: ANCE. Know the gradient-norm argument for why hard negatives matter and why the index must refresh during training.[10]
7. Unsupervised Dense Information Retrieval with Contrastive Learning. Authors: Izacard, Caron, Hosseini et al. Claim to know: Contriever, and how far retrieval can go without labeled pairs.[10]
8. BEIR: A Heterogeneous Benchmark for Zero-shot Evaluation of Information Retrieval Models [core]. Authors: Thakur, Reimers, Ruckle et al. Claim to know: BEIR compares heterogeneous zero-shot tasks and shows strong lexical baselines against many tested dense retrievers. It motivates domain-specific comparison, not a universal dense-versus-lexical ordering.[10]
9. SPLADE: Sparse Lexical and Expansion Model for First Stage Ranking [core]. Authors: Formal, Piwowarski, Clinchant. Claim to know: Study the masked-language-model vocabulary head, activation/pooling and learned expansion. FLOPS regularization is a differentiable sparsity/expected-posting-work proxy, not a measured count of hardware FLOPs or end-to-end latency.[10]
10. ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction [core]. Authors: Khattab and Zaharia. Claim to know: Late interaction and MaxSim. Know the storage cost per token and why that is what ColBERTv2 attacks.[10]
11. ColBERTv2: Effective and Efficient Retrieval via Lightweight Late Interaction. Authors: Santhanam, Khattab, Saad-Falcon et al. Claim to know: Residual compression brings multi-vector storage into range.[10]
12. Passage Re-ranking with BERT. Authors: Nogueira and Cho. Claim to know: Joint query-document scoring is a common second-stage design. Its candidate, length, interaction and batching costs matter; later retrieval systems need not all share this architecture.[10]
13. Document Ranking with a Pre-trained Sequence-to-Sequence Model. Authors: Nogueira, Jiang, Pradeep, Lin. Claim to know: monoT5 renormalizes the logits of the single-token targets true and false: $r=\sigma(z_{\mathrm{true}}-z_{\mathrm{false}})$. It does not rank by the raw true logit. Distinguish this pointwise classification training from pairwise/listwise ranking losses.[11]
14. From RankNet to LambdaRank to LambdaMART: An Overview [core]. Authors: Burges. Claim to know: LambdaRank supplies metric-weighted pseudo-gradients built from pairwise preferences. A RankNet pair loss exists, but rank-dependent weights treated as fixed do not automatically make the whole field the exact gradient of one smooth global ranking objective.[10]
15. Cumulated Gain-Based Evaluation of IR Techniques [core]. Authors: Jarvelin and Kekalainen. Claim to know: DCG combines relevance gains with a chosen rank discount; nDCG divides by ideal DCG for the same judged query and cutoff. Logarithmic discount is a conventional design choice, not a uniquely optimal mathematical consequence.[10]
16. MS MARCO: A Human Generated MAchine Reading COmprehension Dataset. Authors: Nguyen, Rosenberg, Song et al. Claim to know: MS MARCO is widely used for retrieval training/evaluation. Study its annotation, query and candidate biases; it does not represent every workload or establish how most published retrieval numbers were produced.[10]
17. Reciprocal Rank Fusion Outperforms Condorcet and Individual Rank Learning Methods. Authors: Cormack, Clarke, Buettcher. Claim to know: RRF combines ranked lists using a stated smoothing constant, often $c_{\mathrm{RRF}}=60$ in the cited experiments. This is an empirical choice, not a universal optimum or the final context depth.[10]

11. Query understanding and control flow

What. This cluster supports chapter 24 through chapter 26. Why. It connects query reformulation, expansion, retrieval gates, correction, multi-hop control, and learned search. Failure without it. A candidate can propose a controller without its drift mode, generation hallucinations, evaluation cost, supervision source, or stopping rule. Cost and complexity. The cluster contains 16 entries, including 5 marked [core].

1. Relevance feedback in information retrieval. Authors: Rocchio. Claim to know: Study Rocchio’s vector update as a classical relevance-feedback baseline. Modern neural reformulation has related goals but need not literally descend from this exact update.[12]
2. Query Expansion Using Local and Global Document Analysis. Authors: Xu and Croft. Claim to know: Pseudo-relevance feedback, and the drift failure mode that makes unsupervised use risky.[12]
3. Precise Zero-Shot Dense Retrieval without Relevance Labels [core]. Authors: Gao, Ma, Lin, Callan. Claim to know: HyDE generates a hypothetical document and embeds it to retrieve real evidence. The generated document is a query representation, not verified evidence, and can amplify an erroneous assumption.[12]
4. Query2doc: Query Expansion with Large Language Models. Authors: Wang, Yang, Wei. Claim to know: The same idea in expansion form, with a cost comparison worth knowing.[12]
5. Document Expansion by Query Prediction [core]. Authors: Nogueira, Yang, Lin. Claim to know: doc2query moves document-expansion generation work to ingestion and refresh time. Account for corpus churn, generated-query storage and update costs rather than treating expansion as permanently free.[12]
6. Task-Oriented Query Reformulation with Reinforcement Learning. Authors: Nogueira and Cho. Claim to know: Rewriting trained against retrieval reward rather than against a reference rewrite.[12]
7. Active Retrieval Augmented Generation [core]. Authors: Jiang, Xu, Gao et al. Claim to know: FLARE. Know the confidence trigger and the distinction between implicit and explicit query formation.[12]
8. Self-RAG: Learning to Retrieve, Generate, and Critique through Self-Reflection [core]. Authors: Asai, Wu, Wang et al. Claim to know: Reflection tokens as learned control. Know that the critic is distilled into the generator rather than run separately.[12]
9. Corrective Retrieval Augmented Generation. Authors: Yan, Xu, Ran et al. Claim to know: CRAG evaluates retrieved quality, decomposes, and falls back to web search. Know what its evaluator costs and where it is fragile.[12]
10. Adaptive-RAG: Learning to Adapt Retrieval-Augmented LLMs through Question Complexity. Authors: Jeong, Baek, Cho et al. Claim to know: Silver-label supervision for a complexity router, and the honest account of where those labels come from.[12]
11. Unified Active Retrieval for Retrieval Augmented Generation. Authors: Cheng et al. Claim to know: Study the four classifier decisions, frozen-state features, silver labels and their measured routing errors. It is one implementable design, not a proof that every workload needs four classifiers.[12]
12. Interleaving Retrieval with Chain-of-Thought Reasoning [core]. Authors: Trivedi, Balasubramanian, Khot, Sabharwal. Claim to know: IRCoT interleaves retrieval and reasoning to find evidence for later hops. It addresses cases where one-shot retrieval fails; it does not prove that every compositional question requires recursive retrieval.[12]
13. Measuring and Narrowing the Compositionality Gap in Language Models. Authors: Press, Zhang, Min et al. Claim to know: Self-Ask, and a crisp definition of the gap multi-hop retrieval exists to close.[12]
14. Search-R1: Training LLMs to Reason and Leverage Search Engines with Reinforcement Learning. Authors: Jin, Zhang, Yoon et al. Claim to know: Study reinforcement learning over reasoning/search trajectories, including tool protocol, rewards, search budget and termination. Learned stopping is one behavior, not the entire contribution.[12]
15. DeepSeekMath: Pushing the Limits of Mathematical Reasoning. Authors: Shao, Wang, Zhu et al. Claim to know: The GRPO algorithm itself, which Search-R1 optimizes with.[12]
16. HotpotQA: A Dataset for Diverse, Explainable Multi-hop Question Answering. Authors: Yang, Qi, Zhang et al. Claim to know: Inspect the multi-hop labels and shortcut artifacts. Some gains can exploit shortcuts; that does not render every benchmark result illusory.[12]

12. Training the retriever and the generator

What. This cluster supports chapter 27 through chapter 29. Why. It covers pre-training with retrieval, passage fusion, black-box generators, instruction tuning, distractors, differentiability, synthetic data, and supervision limits. Failure without it. A candidate can say “train jointly” without understanding index staleness, decode cost, narrow interfaces, filtering, or recursive-data failure. Cost and complexity. The cluster contains 14 entries, including 6 marked [core].

1. REALM: Retrieval-Augmented Language Model Pre-Training [core]. Authors: Guu, Lee, Tung et al. Claim to know: REALM places retrieval inside pre-training and asynchronously refreshes document representations. Staleness occurs when document-side parameters or corpus/index state change; a query-only update need not invalidate stored document vectors.[13]
2. Leveraging Passage Retrieval with Generative Models for Open Domain Question Answering [core]. Authors: Izacard and Grave. Claim to know: Fusion-in-Decoder. Passages are encoded independently and fused in the decoder. Know what that costs at decode time.[13]
3. FiD-Light: Efficient and Effective Retrieval-Augmented Text Generation. Authors: Hofstatter, Chen, Raman, Zamani. Claim to know: The cost fix for FiD, and the quality it trades.[13]
4. REPLUG: Retrieval-Augmented Black-Box Language Models [core]. Authors: Shi, Min, Yasunaga et al. Claim to know: REPLUG uses a frozen language model and retrieval-conditioned likelihood signals; its retriever learning requires likelihood access. “Black box” does not imply compatibility with every text-only hosted interface.[13]
5. RA-DIT: Retrieval-Augmented Dual Instruction Tuning. Authors: Lin, Chen, Chen et al. Claim to know: Tune both halves without joint backpropagation.[13]
6. RAFT: Adapting Language Model to Domain Specific RAG [core]. Authors: Zhang, Patil, Jain et al. Claim to know: Train with distractors deliberately in the context. This is the most directly actionable training result in this part.[13]
7. The Power of Noise: Redefining Retrieval for RAG Systems. Authors: Cuconasu, Trappolini, Siciliano et al. Claim to know: The counterintuitive finding that some irrelevant context can help, and the conditions under which it does not.[13]
8. Making Retrieval-Augmented Language Models Robust to Irrelevant Context. Authors: Yoran, Wolfson, Ram, Berant. Claim to know: The complementary robustness-training result.[13]
9. Categorical Reparameterization with Gumbel-Softmax. Authors: Jang, Gu, Poole. Claim to know: Gumbel-Softmax provides a continuous relaxation with a temperature trade-off. At finite temperature it is not an exact pathwise gradient of hard categorical or top-k selection, and it is not the only way to learn discrete decisions.[13]
10. InPars: Unsupervised Dataset Generation for Information Retrieval [core]. Authors: Bonifacio, Abonizio, Fadaee, Nogueira. Claim to know: Study synthetic query generation and the filtering/selection recipe together. Filtering matters, but generation, coverage and retriever training also contribute; no universal single-source quality explanation follows.[13]
11. Promptagator: Few-shot Dense Retrieval From 8 Examples [core]. Authors: Dai, Zhao, Ma et al. Claim to know: Promptagator demonstrates task-specific retrieval with eight seed examples in the reported setup. This motivates adaptation, not a proof that a general-purpose retriever can never transfer.[13]
12. Synthetic QA Corpora Generation with Roundtrip Consistency. Authors: Alberti, Andor, Pitler et al. Claim to know: Consistency filtering, stated cleanly.[13]
13. Weak-to-Strong Generalization. Authors: Burns, Izmailov, Kirchner et al. Claim to know: Study when a student trained on weak supervision exceeds its weak teacher and where it fails. A teacher’s aggregate score is not a strict mathematical upper bound on the student.[13]
14. The Curse of Recursion: Training on Generated Data Makes Models Forget. Authors: Shumailov, Shumaylov, Zhao et al. Claim to know: Why a synthetic-data pipeline needs a real-data anchor.[13]

13. Context assembly, attribution, and evaluation

What. This cluster supports chapter 30 through chapter 34. Why. It covers evidence position, context use, attribution, factual precision, retrieval-aware evaluation, inconsistency detection, and judgment-set reliability. Failure without it. A candidate can cite a curve or automatic score without its closed-book baseline, attribution limits, blind spots, or human-agreement gap. Cost and complexity. The cluster contains 17 entries, including 5 marked [core].

1. Lost in the Middle: How Language Models Use Long Contexts [core]. Authors: Liu, Lin, Hewitt et al. Claim to know: Inspect the position-controlled results and matched closed-book baseline. Retrieval-augmented accuracy can fall below closed-book in some evaluated settings; a middle trough is not a universal property of all models and tasks.[14]
2. Make Your LLM Fully Utilize the Context. Authors: An, Ma, Lin et al. Claim to know: IN2 is a training intervention using information-intensive long-context examples. Distinguish training-data/recipe changes from an architectural modification or an inference-only prompt.[14]
3. Found in the Middle: Calibrating Positional Attention Bias. Authors: Hsieh, Chuang, Li et al. Claim to know: Study the proposed positional-attention calibration and its measured effect. Improvement supports that intervention; it does not uniquely prove that all positional failures are mechanical rather than semantic.[14]
4. Differential Transformer. Authors: Ye, Dong, Xia et al. Claim to know: Study the proposed differential-attention mechanism, cost and measured context behavior. Treat it as a tested architectural proposal, not a settled long-run solution.[14]
5. Train Short, Test Long: Attention with Linear Biases. Authors: Press, Smith, Lewis. Claim to know: ALiBi, and the positional-encoding side of the length-extrapolation question.[14]
6. Measuring Attribution in Natural Language Generation Models [core]. Authors: Rashkin, Nikolaev, Lamm et al. Claim to know: AIS separates interpretation from evidence attribution. Resolve the statement’s explicature using its context; a sentence that is not self-contained in isolation can still be interpretable and attributable after context resolution.[14]
7. FEVER: a Large-scale Dataset for Fact Extraction and VERification. Authors: Thorne, Vlachos, Christodoulopoulos, Mittal. Claim to know: The supported / refuted / not-enough-information labels that the attribution taxonomy maps onto.[14]
8. Teaching Language Models to Support Answers with Verified Quotes. Authors: Menick, Trebacz, Mikulik et al. Claim to know: GopherCite, and the cost of doing attribution honestly. Sample, then rerank against the quoted span.[14]
9. WebGPT: Browser-assisted question-answering with human feedback. Authors: Nakano, Hilton, Balaji et al. Claim to know: Retrieval as a browsing policy, with citation as a trained behavior rather than a post-process.[14]
10. AttributionBench: How Hard is Automatic Attribution Evaluation? [core]. Authors: Li, Zhang, Malthus et al. Claim to know: Know how far automatic attribution judges are from human agreement. Cite this whenever someone proposes an LLM judge for grounding.[14]
11. FActScore: Fine-grained Atomic Evaluation of Factual Precision [core]. Authors: Min, Krishna, Lyu et al. Claim to know: Atomic-fact scoring exposes mixed correctness hidden by a coarse answer-level label. Whole-answer adequacy and critical-error measures still provide complementary information.[14]
12. VeriScore: Evaluating the Factuality of Verifiable Claims. Authors: Song, Kim, Iyyer. Claim to know: VeriScore extracts context-aware verifiable claims and combines support precision with $R_K=\min(s/K,1)$, where $s$ counts supported claims and $K$ is a domain cohort median. This is a supported-count proxy, not recall of all missing gold facts or a correction that verifies unverifiable assertions.[15]
13. RAGAs: Automated Evaluation of Retrieval Augmented Generation [core]. Authors: Es, James, Espinosa-Anke, Schockaert. Claim to know: Faithfulness, answer relevance, and context relevance. Know what each does and does not measure before quoting a RAGAS number.[14]
14. SummaC: Re-Visiting NLI-based Models for Inconsistency Detection. Authors: Laban, Schnabel, Bennett, Hearst. Claim to know: Study sentence-level entailment aggregation as a consistency signal, including its evaluated advantages and limits. Failures of a tested document-level NLI setup are not proof that every document-level method must fail.[14]
15. Cumulated Gain-Based Evaluation of IR Techniques. Authors: Jarvelin and Kekalainen. Claim to know: This appears again because the evaluation chapters use it directly. The derivation is in section A.4.[14]
16. How Reliable Are the Results of Large-Scale Information Retrieval Experiments?. Authors: Zobel. Claim to know: Shallow judgment pools and treating unjudged items as irrelevant can bias evaluation. State pooling, coverage and uncertainty; the direction and magnitude are not identical for every system comparison.[14]
17. Overview of the TREC 2019 Deep Learning Track. Authors: Craswell, Mitra, Yilmaz et al. Claim to know: How a serious judgment set is actually built.[14]

14. Trust, robustness, scale, and advanced variants

What. This cluster supports chapter 35 through chapter 41. Why. It covers source reliability, credibility interventions, exposure, diversity, attacks, provenance, routing, attention cost, multimodal retrieval, and graph retrieval. Failure without it. A candidate can promise trustworthy or advanced RAG without the source signal, attack path, authenticity boundary, serving cost, or graph-index objection. Cost and complexity. The cluster contains 18 entries, including 5 marked [core].

1. Retrieval-Augmented Generation with Estimation of Source Reliability [core]. Authors: Hwang, Lee, Park et al. Claim to know: RA-RAG estimates source reliability using cross-source agreement in its stated setup. Agreement can also reflect copying or coordinated errors; this is one scalable estimation approach, not the only possible one.[16]
2. CrAM: Credibility-Aware Attention Modification. Authors: Deng, Zhang, Wang et al. Claim to know: Down-weight attention on heads most influenced by unreliable documents. This is an inference-time intervention with no retraining.[16]
3. Trustworthiness in Retrieval-Augmented Generation Systems: A Survey. Authors: Zhou, Chen, Li et al. Claim to know: The six-aspects-across-three-stages taxonomy used to organize the credibility chapter.[16]
4. Fairness of Exposure in Rankings. Authors: Singh and Joachims. Claim to know: Study stochastic ranking policies and exposure constraints relative to the stated merit/relevance convention. Exposure is the allocated resource; relevance is still needed to define utility and possibly fairness targets.[16]
5. The Use of MMR, Diversity-Based Reranking. Authors: Carbonell and Goldstein. Claim to know: Maximal marginal relevance. This is the classical answer to redundancy in a retrieved set and remains a reasonable baseline.[16]
6. Not What You’ve Signed Up For: Compromising Real-World LLM-Integrated Applications with Indirect Prompt Injection [core]. Authors: Greshake, Abdelnabi, Mishra et al. Claim to know: Study indirect instruction injection through untrusted external content. Retrieval is a common entry path, not the only one; this does not define every RAG security threat.[16]
7. PoisonedRAG: Knowledge Corruption Attacks to Retrieval-Augmented Generation [core]. Authors: Zou, Geng, Wang, Gong. Claim to know: Study the attacker’s corpus access, document budget, success metric and tested models. A small poison budget can be effective in that threat model, not necessarily in every defended deployment.[16]
8. Universal and Transferable Adversarial Attacks on Aligned Language Models. Authors: Zou, Wang, Kolter, Fredrikson. Claim to know: GCG studies optimized adversarial suffixes and transfer under specified access and target assumptions. Report success rates and defenses rather than assuming every guardrail fails identically.[16]
9. Jailbroken: How Does LLM Safety Training Fail?. Authors: Wei, Haghtalab, Steinhardt. Claim to know: Know competing objectives and mismatched generalization as the paper’s explanatory mechanisms, not an exhaustive taxonomy of every possible safety failure.[16]
10. The Instruction Hierarchy: Training LLMs to Prioritize Privileged Instructions. Authors: Wallace, Xiao, Leike et al. Claim to know: Training instruction priority can improve resistance to lower-trust instructions. It is not a formal isolation guarantee and does not replace tool permissions, data controls or attack evaluation.[16]
11. C2PA Technical Specification. Authors: Coalition for Content Provenance and Authenticity. Claim to know: C2PA content credentials bind signed provenance assertions under a verification/trust model. Cryptographic integrity does not certify the truth of every depicted claim; linguistic support/attribution and signed provenance answer different questions.[16]
12. RAGRoute: Learning Which Sources to Query in Federated RAG. Authors: Guerraoui, Kermarrec, Pires et al. Claim to know: Source selection as a learned decision, and the fan-out cost it avoids.[16]
13. FlashAttention: Fast and Memory-Efficient Exact Attention with IO-Awareness. Authors: Dao, Fu, Ermon et al. Claim to know: FlashAttention computes exact attention with reduced high-bandwidth-memory traffic through IO-aware tiling. It does not prove all prefill is bandwidth-bound or remove the quadratic attention arithmetic.[16]
14. Learning Transferable Visual Models From Natural Language Supervision [core]. Authors: Radford, Kim, Hallacy et al. Claim to know: Study CLIP’s contrastive text-image training and the three example bias mechanisms analyzed in Chapter 39. They are not an exhaustive universal list for every multimodal retriever.[16]
15. MRAG-Bench: Vision-Centric Evaluation for Retrieval-Augmented Multimodal Models. Authors: Hu, Shu, Chen et al. Claim to know: What multimodal RAG benchmarks actually measure, and how many are text-answerable.[16]
16. From Local to Global: A Graph RAG Approach to Query-Focused Summarization [core]. Authors: Edge, Trinh, Cheng et al. Claim to know: GraphRAG. Know the community hierarchy, the map-reduce query path, and that index cost is the objection to answer first.[16]
17. LightRAG: Simple and Fast Retrieval-Augmented Generation. Authors: Guo, Xia, Fang, Huang. Claim to know: Compare LightRAG’s entity/relation-oriented retrieval and proposed incremental update procedure with the cited hierarchical-summary design. Refresh costs and quality depend on implementation and workload; lack of that hierarchy alone does not prove an update bound.[16]
18. From Louvain to Leiden: Guaranteeing Well-Connected Communities. Authors: Traag, Waltman, van Eck. Claim to know: Study Leiden’s connectivity guarantees under the stated algorithm/objective and Louvain’s possible disconnected communities. Neither method guarantees a globally optimal partition or semantically correct summaries.[16]

Where to start

What. Follow the source’s ten-paper order when preparation time is one week rather than one month. Why. The route preserves the five decisions, two most important negative results, index arithmetic, and threat model. Failure without it. A compressed plan can reorder the source or omit one of the senior-loop coverage areas. Cost and complexity. The route contains ten papers and gives up the broader coverage of the full curated list. If time allows ten papers, read them in this exact source order:

1. Lewis et al. on RAG.
2. Mallen et al. on when not to trust it.
3. Karpukhin et al. on DPR.
4. Thakur et al. on BEIR.
5. Malkov and Yashunin on HNSW.
6. Jegou et al. on product quantization.
7. Khattab and Zaharia on ColBERT.
8. Liu et al. on Lost in the Middle.
9. Min et al. on FActScore.
10. Greshake et al. on indirect prompt injection. The starter set provides a compact introduction to selected design decisions, negative results, index arithmetic and indirect-injection risks. Its coverage is a study recommendation, not an empirically established description of what most senior interview loops ask.

Diagrams

The source appendix contains zero figures and zero tables. No captioned source visual is recreated or invented here.

Whiteboard pack

What to draw

1. Draw a large rectangle labeled “161 selected readings.”
2. Above it, write “roughly 500 cited works” and draw a narrowing arrow into the rectangle.
3. Divide the rectangle into 14 vertical shelf sections.
4. Label the first shelf “What RAG is” and the last shelf “Trust and advanced variants.”
5. Draw one sample reading card inside a shelf.
6. Write “title,” “listed first authors,” and “claim to know” on the card.
7. Add a [core] seal to the sample card.
8. Write “52 core” beside the seal.
9. Draw a path through ten cards below the shelves.
10. Label that path “one-week starting route.”
11. Add a final note: “Know the result, not only the name.”

Spoken script

Appendix E maps fourteen chapter clusters to 161 entries for 160 distinct titles. Fifty-two entries carry a priority marker. I should explain the mechanism, baseline, assumptions and limits rather than repeat a title or turn one study’s result into a universal law. I distinguish training FLOPs from wall time, PQ combinations from stored scalars, and a supported-count factuality proxy from true semantic recall. The ten-paper route is a starting point, not complete coverage.

Interview traps

Probe 1: Why is recognizing a famous paper title insufficient?

The source expects the specific result, mechanism, comparison, or limit attached to the title. A staff-level interviewer can turn recognition into a follow-up such as asking for the baseline and outcome in Lost in the Middle.

Probe 2: Does [core] mean the other readings can be ignored?

No. The marker means a senior candidate should discuss that entry without prompting, while the full list remains the curated subset worth reading. The 52 marked entries set priority inside 161 readings rather than replacing the other clusters.

Probe 3: Why does the list emphasize negative results?

Negative results reveal conditions under which a design fails, degrades accuracy or does not transfer. They are useful interview preparation, but this appendix does not establish their frequency in secondary sources or interviews, and their study conditions remain part of the result.

Probe 4: Why include papers that are not about retrieval?

Distance concentration, tail latency, and instruction hierarchy still constrain a RAG system. Their mechanisms govern index choice, serving behavior, and the treatment of retrieved text even when the paper itself studies another problem.

Probe 5: What does the ten-paper route preserve when time is short?

It keeps the source’s exact order from foundational RAG through indirect prompt injection. The source says the route covers five decisions, two important negative results, index arithmetic, and the threat model.

Key numbers

00-START-HERE · 01-READING-ROUTES · A4_the_rag_card_and_index_datasheet · A6_notation_quick_reference