Bank/phase 0 foundation - #10
Merged
Merged
Conversation
Verified via git show HEAD:<file> that all tracked files (including 02_interview_bank/01-10, README.md, CONTRIBUTING.md) are already stored as LF in the repository; the CRLF previously observed was a local Windows core.autocrlf checkout artifact only. No renormalization needed. This just prevents autocrlf from ever committing CRLF. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
- scripts/lib/questions.mjs: pure markdown-source parser for the interview-question format (fence-aware Qn heading + <details> block parsing, difficulty tags, exact/near-duplicate title detection, countQuestions() cross-check helper for the build). - scripts/check_questions.mjs: CLI with 'check' (structural validation, --strict promotes the 20-question-per-file target to an error), 'gaps' (per-file progress toward 20 questions / 5-8-7 difficulty mix, with a heuristic rubric-slot breakdown to guide authoring), 'readme' (regenerates README.md's per-file counts, section/grand totals, difficulty mix, badge, and hero mentions from marker comments --write or --check), and 'renumber' (sequential Qn rewrite for one file). - scripts/build_site.mjs: cross-checks source-parsed question count against HTML-extracted quiz items per file, and now fails the build (previously just warned) when a quiz-dir file yields zero questions or the two counts disagree -- catches a question silently dropping out of the quiz due to malformed structure. - README.md: wrapped the per-file tables and totals in HTML comment markers so 'npm run readme' can regenerate them; regenerated to the real current counts (427 architecture + 79 failure-mode + 74 concepts = 580, down from the stale hand-maintained 603). - package.json: check/check:strict/gaps/readme/readme:check/ci scripts. - .github/workflows/deploy.yml: validate before building. - .github/workflows/check.yml (new): runs check+build on every PR with a site-preview artifact. - CONTRIBUTING.md: replaced hand-edit badge/count instructions with the npm script workflow. Verified: 'npm run check' correctly flags the 8 legacy 34-41 files (0 questions, E-LEGACY hits) and every file below the 20-question target as warnings; 'npm run readme' regenerates all counts correctly; 'npm run build' now fails on the 8 unconverted legacy files as intended (fixed by the next commit). Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Converts the legacy prose + bold-Q interview section into the standard ## Qn. ... `[Tag]` + <details> format so it appears on the quiz site (previously silently absent -- 0 quiz items extracted). Folds the existing Threat Model, five Technique walkthroughs, trade-off table, and Compliance Considerations into 20 questions at the 5 Basic / 8 Intermediate / 7 Advanced target mix, converting the 3 existing legacy questions (Q6, Q8 mechanism, Q16 GDPR erasure) in place and adding 17 new questions per the coverage rubric (definition, mechanism, taxonomy comparison, implementation, tuning, eval, failure modes, production cost/latency, residual attack surfaces, system design, research origin, limitations). Adds a Real-World Applications table. Verified: npm run check reports 0 errors, gaps shows 20/20 B5/I8/A7, npm run build no longer errors on this file and the quiz page picks up all 20 questions under the Privacy-Preserving RAG section. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Renames '## Architecture' to the standard '## \xf0\x9f\x8f\x97\xef\xb8\x8f Architecture Flow, Components & Tools' heading. Folds the CDC/Debezium walkthrough, Kafka producer/consumer, micro-batching, freshness SLO metrics, and out-of-order handling code into 20 questions (5B/8I/7A), converting the 3 legacy questions in place and adding 17 new ones per the coverage rubric. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Folds the extraction, linearization/chunking strategies, embedding metadata schema, hybrid retrieval, and arithmetic-generation prompt into 20 questions (5B/8I/7A), converting the 3 legacy questions in place. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Folds the core data structures, full BFS-beam and DFS implementations, cost controller, and ToT vs ReAct vs standard RAG comparison into 20 questions (5B/8I/7A), converting the 3 legacy questions in place. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Renames '## Architecture' to the standard heading. Folds the training objective, hard-negative mining, full training pipeline, offline FAISS indexing, and DPR vs BM25 vs modern embeddings comparison into 20 questions (5B/8I/7A), converting the 3 legacy questions in place. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Moves the architecture section to directly follow the tagline. Folds the action space, two-stage BC+RLHF training pipeline, Toolformer's perplexity-based self-supervision, and Gorilla's verifiable API-calling extension into 20 questions (5B/8I/7A), converting the 3 legacy questions in place. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…estions Renames '## Architecture' to the standard heading. Folds schema-constrained generation, per-field NLI grounding validation, multi-document fan-out synthesis, and the schema-over-fit failure mode into 20 questions (5B/8I/7A), converting the 3 legacy questions in place. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…, 20 questions; fix validator separator rule; regenerate README - Converts the last of the 8 legacy files to the standard format, folding the summary-tree build code, level routing, coarse-to-fine drill-down, and RAPTOR comparison into 20 questions (5B/8I/7A). - scripts/lib/questions.mjs: relaxes the E-HR separator check to tolerate an optional single blockquote annotation line (e.g. a '> Related: ...' cross-reference) between </details> and the --- separator -- a legitimate pre-existing pattern in 07-self-rag.md that the stricter rule was incorrectly flagging as an error. - README.md: regenerated counts now that all 8 legacy files (34-41) contribute 20 questions each instead of 0. Verified: npm run check reports 0 errors across all 61 files; npm run build reports 0 warnings and 0 extraction errors for the first time, with quiz.html now carrying 666 questions (up from 506) -- the 34-41 files are fully live on the site. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, late-interaction fundamentals, comparison to hybrid/RAG-Fusion retrieval, tuning knobs, evaluation methodology, MaxSim's stop-word failure mode, scaling to 100M+ passages, security considerations, cost modeling, a code-search system design scenario, and limitations/evolution. Existing Q1-Q5 unchanged (anchors preserved). Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to LLMLingua and Cache-Augmented Generation, tuning knobs, evaluation, the under-expansion failure mode, scaling via ingestion-time caching, training-investment decision gate, policy miscalibration debugging, cost modeling, a customer-support system design scenario, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering comparison to WebGPT and Deep Research RAG, the pattern's practical origin, citation tracking's architectural role, tuning knobs, evaluation, extraction/relevance-filter failure modes, caching/dedup, a corpus-vs-web routing decision gate, cost modeling, SSRF/redirect security risks, a financial-research system design scenario, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to a single agentic loop and to CoRAG, the pattern's practical origin, tuning knobs, GAIA/DeepSearchQA evaluation, sub-agent context-isolation and dead-end failure modes, the lost-in-the-middle connection to Long-Context RAG, a Deep-Research-vs-single-turn routing gate, cost modeling, multi-agent security risks, a competitive-intelligence system design scenario, budget-monitor debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to RAG-Fusion/HyDE, PEARL's research origin, tuning knobs, evaluation vs static few-shot, near-duplicate/off-pattern retrieval failure modes, library lifecycle maintenance, example ordering effects, a dynamic-vs-static decision gate, cost modeling, poisoned-library security risks, a combined document+example system design scenario, conflicting-convention handling, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Also moved Real-World Applications back to the end of the file after the new questions. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to SURGE's per-field grounding, ALCE's research origin, tuning knobs, NLI-vs-LLM-judge verification tiering, coarse-granularity and unretrievable-claim failure modes, multi-passage attribution, scaling verification cost, a compliance decision gate, adversarial-passage security risks, a legal-research system design scenario, verifier-quality auditing, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to WebGPT and RQ-RAG, the paper's research origin, tuning knobs, evaluation methodology, GRPO's zero-variance-group failure mode, a learned-vs-prompted decision framework, frozen-retriever blind-spot monitoring, a training-investment decision gate, RL rollout cost/infrastructure, security risks specific to RL-trained policies, a biomedical research-assistant system design scenario, gold-label-noise debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to CoRAG and Self-RAG, research origins, tuning knobs, evaluation of retrieval-call efficiency, DeepRAG's tree-search training construction, interpretability differences, an Auto-RAG-vs-DeepRAG selection framework, a per-step-vs-Adaptive-RAG decision gate, training/serving cost profile, security risks specific to per-step policies, an enterprise mixed-freshness system design scenario, parametric-confidence drift detection, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to HyDE, the paper's research origin and UltraDomain benchmark, tuning knobs, evaluation methodology, misleading-clue detection, memory freshness strategy, comparison to Iterative Multi-Hop RAG's reformulation, a memory-model-investment decision gate, cost modeling, corpus-wide poisoning blast-radius security risks, a legal due-diligence system design scenario, compression blind-spot debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to Contextual RAG, the two papers' research origins, tuning knobs, evaluation methodology, semantic-dilution failure mode, hierarchical sub-span extraction, comparison to plain Long-Context RAG, a unit-size decision gate, cost modeling, larger-unit security risks, a legal-contract system design scenario, Self-Route miscalibration debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to Table-Aware RAG, the paper's research origin, tuning knobs, evaluation methodology, the dense-prose failure mode, bounding-box citation grounding, cross-page reasoning, a VisRAG-vs-OCR decision gate, cost modeling, image-specific security risks (visual injection, adversarial images), a financial 10-K hybrid system design scenario, page-rendering-failure debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to KAG, the blog-post research origin, tuning knobs, evaluation against GraphRAG's global search quality, the co-occurrence false-positive failure mode, semantic caching of subgraph summaries, adaptive relevance budgeting, a three-way architecture decision gate, cost/worst-case modeling, co-occurrence-poisoning security risks, an exploratory-archive system design scenario, relevance-test miscalibration debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to Auto-RAG/DeepRAG, the paper's research origin, tuning knobs, three-outcome evaluation methodology, premature-convergence and adaptive-round failure modes, comparison to Verifiable RAG's citation checking, a consolidation-overhead decision gate, multi-call cost modeling, fabrication-laundering and injection security risks, an enterprise frequently-changing-knowledge-base system design scenario, consolidation-bias debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to Search-R1 and RQ-RAG, the paper's research origin, tuning knobs, evaluation methodology, the spurious-correct-chain failure mode, best-of-N reranking implementation, a rejection-sampling-vs-prompting decision gate, training/inference cost modeling, security risks in rejection-sampled training data, a domain-adapted legal/medical system design scenario, reranker-reliability debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 15 questions (Q6-Q20) covering the architecture walkthrough, comparison to CoRAG and Auto-RAG/DeepRAG's action space, the paper's research origin, tuning knobs, action-selection evaluation, distribution-shift failure mode, outcome-filtering implementation, a fine-tuning-investment decision gate, tree-decoding cost modeling, decompose-exploit and perplexity-gaming security risks, a jargon-heavy enterprise search system design scenario, path-selection miscalibration debugging, and limitations. Existing Q1-Q5 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 2: all 15 thin files (30-33, 42-52) are now at 20 questions each. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…42-52) Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 8 questions (Q13-Q20) covering the pipeline-to-failure-mode walkthrough, the 2020 RAG paper's research origin, comparison to Advanced RAG, hyperparameter tuning, vector database selection, top-k mis-tuning failure mode, an HR-chatbot system design scenario, and a closing framework connecting Naive RAG's limitations to the rest of the bank. Existing Q1-Q12 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 8 questions (Q13-Q20) covering RRF's rank-based fusion mechanism, HyDE/reranker research origins, comparison to Modular RAG, why dense retrieval alone misses exact matches, hybrid-search tuning knobs, an ablation-harness implementation, reranking mis-application failure modes, and a closing limitations section pointing toward Modular/Agentic RAG. Existing Q1-Q12 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Adds 8 questions (Q13-Q20) covering the architecture walkthrough, the survey-paper research origin, comparison to Agentic RAG's orchestration and to Advanced RAG's fixed pipeline, pipeline-composition design knobs, routing-accuracy evaluation, inconsistent-module-contract failure mode, and a multi-domain enterprise assistant system design scenario. Existing Q1-Q12 unchanged. Final mix: 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 1 (files 01, 02, 03). Verified: check 0 errors, gaps 20/20 B5/I8/A7. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
… 20 questions each 04-agentic-rag.md (15->20, Q16-20): research origin (ReAct/FLARE), comparison to Adaptive RAG, loop tuning knobs, runaway-loop failure mode, limitations/when-simpler-wins. 11-adaptive-rag.md (13->20, Q14-20): architecture walkthrough, research origin, comparison to Agentic RAG and Corrective RAG, classifier tuning knobs, tier-specific miscalibration failure mode, training-data drift maintenance. 06-corrective-rag.md (12->20, Q13-20): architecture walkthrough, research origin, comparison to Adaptive RAG, distinctive mechanism, the three verdict branches, evaluation methodology, evaluator-bias failure mode, a customer-support system design scenario. All three: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 2 (files 04, 11, 06). Verified: check 0 errors across all three files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…20 questions each 05-graph-rag.md (Q13-20): architecture walkthrough, Microsoft GraphRAG research origin, comparison to LightRAG, distinctive mechanism vs vector RAG, tuning knobs, graph-vs-vector routing, fragmentation/under-connection failure mode, limitations pointing to LazyGraphRAG/LightRAG. 15-lightrag.md (Q13-20): architecture walkthrough, research origin, distinctive mechanism vs GraphRAG's community hierarchy, comparison to HippoRAG, tuning knobs, dual-level evaluation methodology, extraction cost modeling, limitations. 20-hipporag.md (Q13-20): architecture walkthrough, research origin (hippocampal memory theory), comparison to LazyGraphRAG, distinctive mechanism vs GraphRAG, PPR tuning knobs, evaluation vs simple graph traversal, false-synonymy-edge failure mode, limitations. 24-kag.md (Q13-20): architecture walkthrough, Ant Group research origin, comparison to LightRAG, distinctive mechanism vs GraphRAG (logical-form execution), tuning knobs, reasoning-chain evaluation methodology, logical-form misparse failure mode, limitations and right-sizing guidance. All four: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 3 (files 05, 15, 20, 24 -- the graph-based architecture group). Verified: check 0 errors across all four files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ons each 07-self-rag.md (Q13-20): architecture walkthrough, research origin, comparison to Corrective RAG, distinctive single-model mechanism, inference-time tuning knobs, [Retrieve] token calibration, reflection-token disagreement failure mode, a fine-tuning-investment decision-gate design scenario. 08-speculative-rag.md (Q13-20): architecture walkthrough, research origin, comparison to CoRAG's best-of-N, distinctive drafter/verifier mechanism, drafter-quality evaluation, accuracy-vs-latency evaluation, redundant-subset failure mode, limitations. 16-raft.md (Q13-20): training pipeline walkthrough, research origin, comparison to RQ-RAG, distinctive distractor-training mechanism, oracle-absent training data, distractor-robustness evaluation methodology, retriever-swap failure mode, fine-tuning cost/infrastructure overhead. All three: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 4 (files 07, 08, 16). Verified: check 0 errors across all three files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…g.md to 20 questions each 09-multimodal-rag.md (Q13-20): architecture walkthrough, CLIP research origin, comparison to VisRAG, distinctive mechanism, fusion/weighting tuning knobs, cross-modal evaluation, embedding-space misalignment failure mode, limitations. 10-long-context-rag.md (Q13-20): architecture walkthrough, lost-in-the-middle research origin, comparison to LongRAG, distinctive mechanism, tuning knobs, evaluation vs chunked retrieval, corpus-growth ceiling failure mode, limitations. 14-contextual-rag.md (Q13-20): architecture walkthrough, Anthropic blog-post research origin, comparison to HyDE, distinctive mechanism, tuning knobs, ambiguity-segmented evaluation methodology, hallucinated-prefix failure mode, limitations. All three: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 5 (files 09, 10, 14). Verified: check 0 errors across all three files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ag.md to 20 questions each 12-structured-rag.md (Q13-20): architecture walkthrough, Spider/DIN-SQL/BIRD research origin, comparison to Table-Aware RAG, distinctive mechanism, tuning knobs, execution-accuracy evaluation, semantically-wrong-SQL failure mode, limitations vs semantic-layer alternatives. 13-raptor.md (Q13-20): architecture walkthrough, research origin, comparison to Recursive Document Summarization RAG, distinctive GMM soft-clustering mechanism, clustering implementation, cluster-boundary-misalignment failure mode, limitations, a tree-depth decision framework. 21-memory-conversational-rag.md (Q13-20): architecture walkthrough, MemGPT research origin, disambiguation from MemoRAG's unrelated use of 'memory', distinctive statefulness mechanism, memory tuning knobs, topic-boundary detection, memory-loss failure mode, systematic multi-turn testing methodology. All three: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 6 (files 12, 13, 21). Verified: check 0 errors across all three files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…g.md, 23-flare-rag.md to 20 questions each 17-cache-augmented-generation.md (Q13-20): architecture walkthrough, research origin, comparison to REFRAG, distinctive no-retrieval mechanism, tuning knobs, evaluation vs RAG, cache-budget-exceeded failure mode, corpus-subset-selection guidance. 18-rag-fusion.md (Q13-20): architecture walkthrough, blog-post research origin, comparison to CoRAG's sequential chain, distinctive relationship to Multi-Query retrieval, reformulation tuning knobs, diversity-vs-redundancy evaluation, query-drift failure mode, limitations. 22-hyde-rag.md (Q13-20): architecture walkthrough, research origin, comparison to Contextual Retrieval, distinctive pseudo-document mechanism, tuning knobs, topic-drift-hallucination failure mode, an LLM-call-investment decision gate, cost modeling. 23-flare-rag.md (Q13-20): architecture walkthrough, research origin, comparison to Self-RAG's [Retrieve] token, distinctive mid-generation-trigger mechanism, tuning knobs, per-sentence-position evaluation methodology, confidence-miscalibration failure mode, scaling retrieval-trigger cost. All four: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 7 (files 17, 18, 22, 23). Verified: check 0 errors across all four files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…alm.md to 20 questions each 19-iterative-multihop-rag.md (Q13-20): architecture walkthrough, IRCoT/Self-Ask research origin, comparison to HippoRAG's single-pass retrieval, distinctive conditioned-query mechanism, loop tuning knobs, per-hop contribution evaluation, wrong-intermediate-fact failure mode, limitations pointing to HippoRAG/CoRAG/Search-R1. 25-graphreader-gnn-rag.md (Q13-20): architecture walkthrough, research origin, comparison to HippoRAG, distinctive trained-GNN mechanism vs GraphRAG's community detection, tuning knobs, evaluation vs PPR heuristics, training-production graph mismatch failure mode, limitations. 26-realm.md (Q13-20): architecture walkthrough, research origin, comparison to DPR, why joint end-to-end training was harder than two-stage approaches, tuning knobs, index-drift detection, a joint-training-vs-frozen-retriever decision framework, REALM's influence on modern frozen-retriever RAG design. All three: existing questions unchanged, final mix 5 Basic / 8 Intermediate / 7 Advanced. This completes Phase 3 batch 8 (files 19, 25, 26). Verified: check 0 errors across all three files, gaps 20/20 B5/I8/A7 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
…ons each 27-retro.md (Q13-20): architecture walkthrough, research origin, comparison to DPR/REALM training approaches, distinctive frozen-retriever-plus-CCA mechanism, tuning knobs, retriever-vs-CCA bottleneck diagnosis, datastore-training-overlap contamination failure mode, datastore-scale relationship to the trillion-token claim. 28-atlas.md (Q13-20): architecture walkthrough, research origin, comparison to Fusion-in-Decoder, distinctive few-shot-joint-training mechanism vs REALM, tuning knobs, few-shot evaluation methodology, joint-training instability failure mode, relationship to modern in-context learning. 29-fusion-in-decoder.md (Q13-20): architecture walkthrough, research origin, comparison to RETRO's chunked cross-attention, distinctive decoder-fusion mechanism, tuning knobs, independent-vs-joint-encoding evaluation, contradictory-passages failure mode, lasting influence on modern RAG readers. All three: existing questions unchanged, final mix 5 Basic / 7-8 Intermediate / 7-8 Advanced (within tolerance). This completes Phase 3 and the full campaign: all 52 architecture files in 02_interview_bank/ now have exactly 20 questions each (1040 total), up from the original 427 across an uneven 3-20 range, plus the 8 legacy prose-format files fully converted to the standard format. Verified: check 0 errors across all three files, gaps 20/20 each. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
Final counts: 1040 architecture questions (52 files x 20) + 79 failure-mode questions + 74 concept questions = 1193 total, up from the stale hand-maintained 603 badge value at the start of this work. Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>
This file contains hidden or bidirectional Unicode text that may be interpreted or compiled differently than what appears below. To review, open the file in an editor that reveals hidden Unicode characters.
Learn more about bidirectional Unicode characters
Sign up for free
to join this conversation on GitHub.
Already have an account?
Sign in to comment
1 participant
Add this suggestion to a batch that can be applied as a single commit.This suggestion is invalid because no changes were made to the code.Suggestions cannot be applied while the pull request is closed.Suggestions cannot be applied while viewing a subset of changes.Only one suggestion per line can be applied in a batch.Add this suggestion to a batch that can be applied as a single commit.Applying suggestions on deleted lines is not supported.You must change the existing code in this line in order to create a valid suggestion.Outdated suggestions cannot be applied.This suggestion has been applied or marked resolved.Suggestions cannot be applied from pending reviews.Suggestions cannot be applied on multi-line comments.Suggestions cannot be applied while the pull request is queued to merge.Suggestion cannot be applied right now. Please check back later.
What does this PR add or change?
Type of change
Checklist
<details>answer block)