{"sources":[{"id":"S01","url":"https://arxiv.org/html/2005.11401v4","title":"Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks","publisher":"arXiv; NeurIPS 2020","published_at":"2020-05-22","accessed_at":"2026-11-16","version":"arXiv v4, 2021-04-12","document_id":"","revision_id":"","limitations":"Historical model formulation; not proof of modern context-stuffing behavior or factual guarantees."},{"id":"S02","url":"https://arxiv.org/pdf/1603.09320","title":"Efficient and robust approximate nearest neighbor search using Hierarchical Navigable Small World graphs","publisher":"arXiv","published_at":"2016-03-30","accessed_at":"2026-11-16","version":"v4, 2018-08-14; version confirmed on arXiv abstract page","document_id":"","revision_id":"","limitations":"Conditional complexity analysis and reported experiments; no universal latency guarantee. HTML unavailable; PDF full text read."},{"id":"S03","url":"https://raw.githubusercontent.com/nmslib/hnswlib/master/ALGO_PARAMS.md","title":"HNSW algorithm parameters","publisher":"nmslib/hnswlib maintainers","published_at":null,"accessed_at":"2026-11-16","version":"master as accessed; commit not pinned","document_id":"","revision_id":"","limitations":"Mutable implementation documentation; parameter guidance is not universal."},{"id":"S04","url":"https://www.staff.city.ac.uk/~sbrp622/papers/foundations_bm25_review.pdf","title":"The Probabilistic Relevance Framework: BM25 and Beyond","publisher":"Foundations and Trends in Information Retrieval; author-hosted copy","published_at":"2009","accessed_at":"2026-11-16","version":"Volume 3(4), pages 333–389; DOI 10.1561/1500000019","document_id":"","revision_id":"","limitations":"BM25 variants differ; no library-specific defaults inferred."},{"id":"S05","url":"https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf","title":"Reciprocal Rank Fusion outperforms Condorcet and individual Rank Learning Methods","publisher":"ACM SIGIR 2009; author-hosted copy","published_at":"2009","accessed_at":"2026-11-16","version":"SIGIR 2009, pages 758–759","document_id":"","revision_id":"","limitations":"Experimental choice k=60; finite-list missing-item and tie conventions are explicitly defined by this lab."},{"id":"S06","url":"https://arxiv.org/html/1901.04085v3","title":"Passage Re-ranking with BERT","publisher":"arXiv","published_at":"2019-01-30","accessed_at":"2026-11-16","version":"v3","document_id":"","revision_id":"","limitations":"Specific BERT reranker and datasets; no universal superiority or latency claim."},{"id":"S07","url":"https://arxiv.org/html/2004.12832v2","title":"ColBERT: Efficient and Effective Passage Search via Contextualized Late Interaction over BERT","publisher":"ACM SIGIR 2020; arXiv","published_at":"2020-06-04","accessed_at":"2026-11-16","version":"v2","document_id":"","revision_id":"","limitations":"Original ColBERT architecture; no claim that every later multi-vector model is identical."},{"id":"S08","url":"https://huggingface.co/blog/multi-vector-encoder","title":"Multi-Vector (Late Interaction) Embedding Models with Sentence Transformers","publisher":"Hugging Face / Sentence Transformers","published_at":"2026-08-18","accessed_at":"2026-11-16","version":"Public saved revision and live page as accessed","document_id":"4c879f7e590d798f7ba989555e6ce9ca","revision_id":"77d567d00a3ce8e7620b4ad9ba01738a","limitations":"Used only for representation and storage explanation; reported timings, package versions, and source code are not adopted."},{"id":"S09","url":"https://arxiv.org/html/2307.03172v3","title":"Lost in the Middle: How Language Models Use Long Contexts","publisher":"arXiv","published_at":"2023-11-20","accessed_at":"2026-11-16","version":"v3","document_id":"","revision_id":"","limitations":"Results apply to tested models and tasks; no universal context-ordering prescription."},{"id":"S10","url":"https://aclanthology.org/2024.eacl-demo.16.pdf","title":"RAGAS: Automated Evaluation of Retrieval Augmented Generation","publisher":"Association for Computational Linguistics","published_at":"2024-03","accessed_at":"2026-11-16","version":"EACL 2024 System Demonstrations, pages 150–158","document_id":"","revision_id":"","limitations":"Original three metrics; LLM-dependent estimates, not factual certification."},{"id":"S11","url":"https://docs.ragas.io/en/stable/concepts/metrics/available_metrics/context_precision/","title":"Context Precision","publisher":"Ragas maintainers","published_at":null,"accessed_at":"2026-11-16","version":"stable documentation as accessed; package version not pinned","document_id":"","revision_id":"","limitations":"Mutable API documentation; used for metric semantics only."},{"id":"S12","url":"https://docs.ragas.io/en/stable/concepts/metrics/available_metrics/context_recall/","title":"Context Recall","publisher":"Ragas maintainers","published_at":null,"accessed_at":"2026-11-16","version":"stable documentation as accessed; package version not pinned","document_id":"","revision_id":"","limitations":"Reference-dependent metric variants; no source-supplied code executed."},{"id":"S13","url":"https://www.microsoft.com/en-us/research/wp-content/uploads/2016/02/MSR-TR-2010-82.pdf","title":"From RankNet to LambdaRank to LambdaMART: An Overview","publisher":"Microsoft Research","published_at":"2010","accessed_at":"2026-11-16","version":"MSR-TR-2010-82","document_id":"","revision_id":"","limitations":"Used for exponential-gain nDCG definition, not the invented grades or results."},{"id":"S14","url":"https://arxiv.org/html/2210.03629v3","title":"ReAct: Synergizing Reasoning and Acting in Language Models","publisher":"arXiv; ICLR 2023","published_at":"2023-03-10","accessed_at":"2026-11-16","version":"v3","document_id":"","revision_id":"","limitations":"Original interaction paradigm; does not define MCP, JSON Schema, or production authorization."},{"id":"S15","url":"https://modelcontextprotocol.io/specification/2025-06-18/architecture","title":"Architecture","publisher":"Model Context Protocol maintainers","published_at":"2025-06-18","accessed_at":"2026-11-16","version":"Specification 2025-06-18","document_id":"","revision_id":"","limitations":"Pinned revision, not a claim about the newest specification or deployed compliance."},{"id":"S16","url":"https://modelcontextprotocol.io/specification/2025-06-18/server/tools","title":"Tools","publisher":"Model Context Protocol maintainers","published_at":"2025-06-18","accessed_at":"2026-11-16","version":"Specification 2025-06-18","document_id":"","revision_id":"","limitations":"Normative requirements and recommendations do not prove implementation enforcement."},{"id":"S17","url":"https://modelcontextprotocol.io/specification/2025-06-18/server/resources","title":"Resources","publisher":"Model Context Protocol maintainers","published_at":"2025-06-18","accessed_at":"2026-11-16","version":"Specification 2025-06-18","document_id":"","revision_id":"","limitations":"Intended application-driven interaction permits UI flexibility."},{"id":"S18","url":"https://modelcontextprotocol.io/specification/2025-06-18/server/prompts","title":"Prompts","publisher":"Model Context Protocol maintainers","published_at":"2025-06-18","accessed_at":"2026-11-16","version":"Specification 2025-06-18","document_id":"","revision_id":"","limitations":"User-controlled is an intended interaction pattern, not a mandatory UI."}],"claims":[{"id":"C01","text":"RAG distinguishes generator parameters from an external index; RAG-Sequence marginalizes document-conditioned sequence probabilities, RAG-Token marginalizes per token, and DPR uses independent query/document encoding with exponential inner-product scoring.","location":"memory paragraphs 1–3 and equations; context architecture table","evidence":[{"source_id":"S01","passage":"the model uses the same document to predict each target token","location":"§2 Methods; §§2.1–2.3 and displayed equations"},{"source_id":"S01","passage":"DPR follows a bi-encoder architecture","location":"§2.2 Retriever: DPR"}],"limitations":"Historical RAG formulation; context concatenation in this example is explicitly a different design.","kind":"source","basis":"Full-text equations and definitions read; no benchmark generalization."},{"id":"C02","text":"The notebook proposes a context-assembly architecture with revisioned chunks, authorization before both retrieval branches, RRF followed by reranking, and separate relevance, support, and source-quality review. Diagram arrows express that original design and do not guarantee truth.","location":"memory design and scope; hybrid diagram and explanation; context experiment; evaluation methodology; hybrid.drawio","evidence":[],"limitations":"Original design prescription, not a measured implementation or universal architecture.","kind":"original_example","basis":"Inspectable hybrid.drawio nodes q,p,b,d,f,r,g and edges e0–e6; explicit design assumptions and scope paragraphs."},{"id":"C03","text":"HNSW uses randomized nested graph layers, greedy upper-layer routing and broader base-layer search; M and layer degree caps are distinct, and logarithmic complexity analysis is conditional.","location":"hnsw paragraphs 1–2 and parameter table; Q2","evidence":[{"source_id":"S02","passage":"an exponentially decaying probability distribution","location":"§4 Algorithm Description, PDF page 4; Algorithms 1 and 5"},{"source_id":"S02","passage":"under the assumption that we build exact Delaunay graphs","location":"§4.2.1 Search complexity, PDF page 7"}],"limitations":"Approximate graphs and arbitrary high-dimensional datasets do not inherit an unconditional worst-case guarantee.","kind":"source","basis":"Algorithms 1–5 and complexity qualifications read in the PDF."},{"id":"C04","text":"hnswlib documents construction/search tradeoffs: M affects links and memory, efConstruction controls construction effort and quality, and query ef increases search effort and generally accuracy; ef must be at least requested k.","location":"hnsw parameter table and implementation paragraph; Q8","evidence":[{"source_id":"S03","passage":"leads to more accurate but slower search","location":"Search parameters, lines 2–4"},{"source_id":"S03","passage":"Bigger ef_construction leads to longer construction, but better index quality.","location":"Construction parameters, lines 20–22"}],"limitations":"Implementation guidance, not a guarantee of monotonic improvement for every query or a portable default.","kind":"source","basis":"Complete short ALGO_PARAMS.md read; mutable master identified."},{"id":"C05","text":"The exhaustive dot-product baseline has O(Nd) arithmetic; ANN recall differs from semantic relevance. AX-17 examples, the 2500-token budget, and the context-support-versus-truth counterexample are synthetic deductions.","location":"hnsw baseline; hybrid identifier example; context budget and experiment; evaluation counterexample; Q5 and Q8","evidence":[],"limitations":"Assumed dense dot products and invented token allowance; excludes hardware costs, real tokenization, and empirical model behavior.","kind":"original_example","basis":"N repetitions of a d-component dot product; 4096−600−96−800−100=2500; explicitly false invented passage repeated faithfully provides the logical counterexample."},{"id":"C06","text":"BM25 sums term contributions using saturation and length normalization; parameter choices and formula variants differ, including scaling conventions.","location":"hybrid BM25 equation and definitions","evidence":[{"source_id":"S04","passage":"the full document score is obtained by summing these term-weights","location":"§3.4, equation 3.15, printed page 360"},{"source_id":"S04","passage":"Published versions of BM25 can vary somewhat","location":"§3.5.1, printed page 361"}],"limitations":"The displayed common form includes k1+1 scaling; no universal score sign or bound is asserted.","kind":"source","basis":"Equations 3.12–3.15 and variations discussion read."},{"id":"C07","text":"RRF sums reciprocal smoothed ranks rather than raw scores; the paper fixed k=60 after a pilot. Rank-only fusion loses score-margin information and this lab defines missing-item contributions as zero.","location":"hybrid RRF explanation; Q1; lab RRF premise","evidence":[{"source_id":"S05","passage":"where k = 60 was fixed during a pilot investigation","location":"§1, page 1, immediately after RRF formula"},{"source_id":"S05","passage":"without regard to the arbitrary scores returned by particular ranking methods","location":"§2 Discussion, page 2"}],"limitations":"Truncated-list, duplicate, and tie policies are lab choices; no universal superiority claimed.","kind":"source","basis":"Both pages read, including scoring equation and pilot/validation distinction."},{"id":"C08","text":"The BERT cross-encoder jointly scores query and passage candidates, motivating shortlisted reranking; pair evaluation count is distinct from batch launch count.","location":"context table and reranker paragraph; Q3","evidence":[{"source_id":"S06","passage":"we feed the query as sentence A and the passage text as sentence B","location":"§2 Method"},{"source_id":"S06","passage":"for each passage independently","location":"§2 Method, relevance probability computation"}],"limitations":"Batching distinction is an inference from independent pair evaluation, not a measured latency result.","kind":"source","basis":"Task and Method sections read; no top-50 mandate or broad superiority extrapolated."},{"id":"C09","text":"ColBERT independently encodes query/document tokens and sums maximum document-token similarities for each query token, permitting offline document encoding.","location":"context architecture table and ColBERT paragraph","evidence":[{"source_id":"S07","passage":"maximum similarity (e.g., cosine similarity), and the scalar outputs of these operators are summed across query terms","location":"Introduction architecture description; §3.3 Late Interaction"}],"limitations":"Original ColBERT mechanism; not all late-interaction variants.","kind":"source","basis":"Full-text architecture and MaxSim sections read."},{"id":"C10","text":"Multi-vector representations retain token vectors rather than a single document vector, creating storage tradeoffs that depend on representation and compression.","location":"context table and ColBERT storage discussion","evidence":[{"source_id":"S08","passage":"One vector per token instead of one vector per document is a lot more vectors","location":"What You Gain, and What It Costs; saved text beginning after offset 3500"}],"limitations":"No borrowed timings, storage ratios, model compatibility, or package execution claims.","kind":"source","basis":"ResearchNode read_document offsets 0 and 3500 for exact saved document/revision; live public page opened."},{"id":"C11","text":"Lost in the Middle reports position-sensitive performance, often better at the start/end than middle, for tested tasks and models; it motivates testing rather than a universal ordering rule.","location":"context position paragraph; Q4","evidence":[{"source_id":"S09","passage":"performance is often highest when relevant information occurs at the beginning or end of the input context","location":"Abstract; Figure 1 and §2.3"}],"limitations":"Historical empirical finding; no claim that all current models behave identically.","kind":"source","basis":"Full-text experimental framing and findings read."},{"id":"C12","text":"Original RAGAS uses faithfulness, answer relevance, and context relevance proxies; support fraction and generated-question similarity depend on model judgments and are not factual certification.","location":"evaluation original metrics; Q5","evidence":[{"source_id":"S10","passage":"the number of statements that were supported according to the LLM","location":"§3, faithfulness formula, PDF page 3"},{"source_id":"S10","passage":"It relies heavily on the performance of the LLMs","location":"§8 Limitations, PDF page 7"}],"limitations":"Original paper’s metric definitions are distinguished from current documentation; no independent expert verification.","kind":"source","basis":"§3 formulas for faithfulness, answer relevance, context relevance and §8 read."},{"id":"C13","text":"Ragas context precision is rank-sensitive and the current reference-based variant compares retrieved contexts against a reference answer.","location":"evaluation current metrics paragraph","evidence":[{"source_id":"S11","passage":"rank relevant chunks higher than irrelevant ones","location":"Context Precision definition, line 254; displayed formula and Examples"}],"limitations":"Mutable stable documentation; no API execution or version compatibility claim.","kind":"source","basis":"Definition, formula, and distinction from Context Utilization read."},{"id":"C14","text":"Ragas context recall needs a reference; its LLM variant checks attribution of reference-answer claims to retrieved context.","location":"evaluation current metrics paragraph","evidence":[{"source_id":"S12","passage":"calculating context recall always requires a reference to compare against","location":"Context Recall definition, line 240; following claim-attribution formula"}],"limitations":"Depends on reference coverage and evaluator; not universally reference-free.","kind":"source","basis":"Current definition and formula read."},{"id":"C15","text":"Exponential-gain nDCG normalizes discounted graded relevance by maximum attainable DCG for the query, using the same cutoff and judgment universe.","location":"evaluation nDCG formula; lab denominator explanation; lab.py dcg and ndcg","evidence":[{"source_id":"S13","passage":"the denominator is the maximum DCG@T attainable for that query","location":"§3 Information Retrieval Measures, equation 5 and normalization, PDF page 5"}],"limitations":"Lab uses base-two logarithms and a complete synthetic judgment universe; zero ideal gain convention is original.","kind":"source","basis":"Equation 5 and normalized form read; changing logarithm base consistently cancels in the ratio."},{"id":"C16","text":"ReAct interleaves reasoning, actions, and observations; its original action framework does not require MCP or JSON Schema and does not guarantee self-correction.","location":"tools opening; Q6","evidence":[{"source_id":"S14","passage":"an agent receives an observation","location":"§2 ReAct formal environment setup"},{"source_id":"S14","passage":"reasoning traces and task-specific actions in an interleaved manner","location":"Abstract; §2"}],"limitations":"Mechanism retained; comparative benchmark gains not generalized.","kind":"source","basis":"Original full-text interaction setup read."},{"id":"C17","text":"MCP revision 2025-06-18 separates host/client/server roles; hosts enforce security and authorization, and capabilities identify supported protocol features rather than user permissions.","location":"tools pinned architecture paragraph; Q7","evidence":[{"source_id":"S15","passage":"Handles user authorization decisions","location":"Core Components: Host"},{"source_id":"S15","passage":"Capabilities determine which protocol features and primitives are available during a session","location":"Capability Negotiation"}],"limitations":"Pinned historical revision; permission/capability distinction is an inference from their separately specified responsibilities.","kind":"source","basis":"Architecture host duties and capability negotiation read."},{"id":"C18","text":"MCP tools have names and schemas, are intended to be model-controlled with flexible UI, and separately require validation and access controls; human oversight is recommended.","location":"tools table and controls paragraph; Q7","evidence":[{"source_id":"S16","passage":"designed to be model-controlled","location":"User Interaction Model"},{"source_id":"S16","passage":"Validate all tool inputs","location":"Security Considerations: Servers MUST"},{"source_id":"S16","passage":"Implement proper access controls","location":"Security Considerations: Servers MUST"},{"source_id":"S16","passage":"human in the loop","location":"User Interaction Model, SHOULD recommendation"}],"limitations":"SHOULD is not MUST; schema validation does not prove authorization or safe effects.","kind":"source","basis":"Input schema example, interaction model, and security requirements read."},{"id":"C19","text":"MCP resources expose URI-identified context and are intended to be application-driven, with flexible interaction patterns.","location":"tools table and control paragraph; Q7","evidence":[{"source_id":"S17","passage":"Each resource is uniquely identified by a URI","location":"Resources introduction"},{"source_id":"S17","passage":"designed to be application-driven","location":"User Interaction Model"}],"limitations":"No blanket assertion that resources are static files or enforce data permissions themselves.","kind":"source","basis":"Introduction and interaction model read."},{"id":"C20","text":"MCP prompts are templates intended for explicit user selection, while implementations may vary their UI.","location":"tools table and control paragraph; Q7","evidence":[{"source_id":"S18","passage":"Prompts are designed to be user-controlled","location":"User Interaction Model"},{"source_id":"S18","passage":"the protocol itself does not mandate any specific user interaction model","location":"User Interaction Model"}],"limitations":"Control label is not a mandatory interface or authorization mechanism.","kind":"source","basis":"Prompt definition and user interaction paragraphs read."},{"id":"C21","text":"The original guarded-loop design separates proposed actions, argument checks, host authorization/budget, execution, untrusted observations, finish and failure; the local gate does not implement a complete sandbox or MCP stack.","location":"tools diagram, text alternative and arrow explanations; host design paragraph; lab tool exercise; Q6; tools.drawio; lab.py guarded_call","evidence":[],"limitations":"Trusted principal and attempt counter are assumed host inputs; no real service or security enforcement tested.","kind":"original_example","basis":"tools.drawio nodes m,v,p,e,o,s,f and labeled edges; lab.py PERMISSIONS, guarded_call, local synthetic observation and remaining-attempt accounting."},{"id":"C22","text":"Synthetic RRF table/order, nDCG values, exact-tie policy, edge-case assertions, and narrow tool validation behavior follow from the original lab and stated fixtures; all code is illustrative.","location":"entire lab section, code and table; Q1; lab.py including comments, schema, and assertions","evidence":[],"limitations":"Python not executed at author stage; arithmetic was separately evaluated in JavaScript. No measured retrieval quality, latency, schema-library compliance, or security guarantee.","kind":"original_example","basis":"lab.py: dense C,A,D,B; lexical B,A,E,C; grades A=3,B=2,C=2,D=0,E=1; c=60; Fraction accumulation; alphabetic ties; full judgments for ideal DCG. Separate arithmetic tool result: 2/62=0.03225806451612903, 1/61+1/64=0.032018442622950824, 1/63=0.015873015873015872; dense nDCG@3=0.7136205775898136; fused=1; five-grade ratio=0.8322420383257692. validate_call and guarded_call are inspectable definitions, not executed evidence."},{"id":"C23","text":"Source disclosure accurately describes public-source reading, saved-original identity, explicit specification versions, and separation of synthetic arithmetic from empirical evidence.","location":"source-checks; bibliography metadata and scope","evidence":[],"limitations":"No claim of publication approval, Python execution, browser testing, or independent expert review.","kind":"process","basis":"Tool record: focused ResearchNode HNSW, RAG, ReAct and reciprocal rank fusion searches; read_document for S08 at offsets 0 and 3500; web opens of listed primary full texts and relevant sections; draw.io XML reference followed by create_diagram calls for both assets; separate JavaScript arithmetic. Attached reference images guide structure; no screenshot acceptance claimed."}]}