# Primary sources and scope ledger

Checked by browsing on **2026-10-17**. These references support concepts; they are not copied course material. Examples, synthetic records, exercise wording and the diagram were authored for this draft. No benchmark result from a paper is generalized to this project.

| Source | Used here | Boundaries |
| --- | --- | --- |
| Lewis et al., [Retrieval-Augmented Generation for Knowledge-Intensive NLP Tasks](https://arxiv.org/abs/2005.11401), NeurIPS 2020, revised 2021 | Module 1 explains combining retrieved external material with generation | The paper's trained dense-retrieval/sequence-to-sequence architecture is not implemented by this lexical demo |
| Manning, Raghavan and Schütze, [TF-IDF weighting](https://nlp.stanford.edu/IR-book/html/htmledition/tf-idf-weighting-1.html), *Introduction to Information Retrieval* | The project's lexical baseline and authored Module4 | The project states its own smoothing and raw-count term-frequency formula; this is one variant, not the book's formula verbatim |
| Same authors, [Dot products](https://nlp.stanford.edu/IR-book/html/htmledition/dot-products-1.html) | Cosine normalization and the limits of bag-of-words matching | Word order and negation are not understood by the index |
| Same authors, [Evaluation of unranked retrieval sets](https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-unranked-retrieval-sets-1.html) and [Evaluation of ranked retrieval results](https://nlp.stanford.edu/IR-book/html/htmledition/evaluation-of-ranked-retrieval-results-1.html) | Recall and evaluation at a fixed ranking cutoff | The project defines reciprocal rank directly; tiny teaching fixtures are not representative user studies |
| NIST, [Generative Artificial Intelligence Profile, AI 600-1](https://www.nist.gov/publications/artificial-intelligence-risk-management-framework-generative-artificial-intelligence), 2024 | Module9 risk and bounded-readiness framing | A voluntary risk-management reference, not a compliance claim or deployment authorization |
| Liu et al., [Lost in the Middle: How Language Models Use Long Contexts](https://arxiv.org/abs/2307.03172), 2023/2024 | Module6 design of a context-position investigation | Findings depend on evaluated models/tasks; this course has no long-context neural experiment |
| Python, [unittest documentation](https://docs.python.org/3/library/unittest.html) | Standard-library test discovery and assertions | Tested interpreter is Python 3.14.4; the unversioned documentation currently points to a newer stable version |
| Jurafsky and Martin, [N-gram Language Models](https://web.stanford.edu/~jurafsky/slp3/3.pdf), *Speech and Language Processing*, draft 2026-08-19 | Module 2 language-model terminology and count-based training/inference | All toy sentences, counts and decoding traces are original; this is not a neural model |
| Vaswani et al., [Attention Is All You Need](https://arxiv.org/abs/1706.03762), 2017, arXiv revision 2023 | Module 2 Transformer/attention distinction | Scalar weighted-average arithmetic is not a trained Transformer or attention explanation for real outputs |
| W3C, [PROV Overview](https://www.w3.org/TR/prov-overview/), Working Group Note 2013 | Module 3 provenance/derivation context | Parent spans are a small local design; the course does not implement the PROV family |
| Karpukhin et al., [Dense Passage Retrieval for Open-Domain Question Answering](https://arxiv.org/abs/2004.04906), 2020 | Module 5 learned dual-encoder concept | The runnable vectors are manually chosen simulations; no DPR benchmark or learned encoder is reproduced |
| Cormack, Clarke and Büttcher, [Reciprocal Rank Fusion](https://plg.uwaterloo.ca/~gvcormac/cormacksigir09-rrf.pdf), SIGIR 2009 | Module 5 rank-fusion formula | The formula runs on original tiny lists; its score is not confidence, and improved quality is not presumed |
| Nogueira and Cho, [Passage Re-ranking with BERT](https://arxiv.org/abs/1901.04085), 2019 | Module 5 distinction between candidate reranking and retrieval | The executed exact-term constraint is not a neural reranker |
| Shi et al., [Large Language Models Can Be Easily Distracted by Irrelevant Context](https://arxiv.org/abs/2302.00093), 2023 | Module6 motivation for testing distractors separately | No distraction/position result is claimed for an unrun model |
| Greshake et al., [Not what you've signed up for: Indirect Prompt Injection](https://arxiv.org/abs/2302.12173), 2023 | Module9 source-data/instruction distinction | Inert local text and fingerprint checks are not neural injection-resistance or production security validation |
| Min et al., [FActScore](https://arxiv.org/abs/2305.14251), 2023 | Module8 motivation for claim-level support review | The evaluator aggregates manual labels; it does not reproduce FActScore's automated method |
| Saad-Falcon et al., [ARES](https://arxiv.org/abs/2311.09476), 2023/2024 | Module8 automated-evaluation research context | No learned judge runs; proposed judges need separate validation |
| Es et al., [Ragas](https://arxiv.org/abs/2309.15217), 2023/2024 | Module8 component-evaluation research context | No Ragas installation, current-library API claim or automatic score is used |

## Repository duplication check

Read-only inspection of `/Users/anandpaiv/Workspace/Projects/AabhaAI/content/courses` on 2026-10-17 found eleven `course.json` packages, plus FastAPI launch/v1/v2 manifests. Their titles and module outlines were reviewed. Filename and Markdown/JSON text searches for Generative AI, retrieval-augmented and LLMOps did not reveal an equivalent advanced course draft. Text matches were confined to Everyday AI material/revisions. This is a scoped repository check, not a search of all private drives or conversations.

Relevant adjacent courses:

- **AI for Everyday Work and Prompting:** works from supplied documents and checks claims/evidence.
- **Codeless AI Assistants and Automation:** builds a trustworthy knowledge pack and operates a reviewed workflow.
- **Deep Learning Foundations:** introduces attention as one bridge beyond the core.
- **FastAPI launch/v2:** builds an Equipment Rental API. No rental API lessons, web server scaffolding or API deployment work are reused here.

The parent implementation task also reported no equivalent advanced draft in its prior release/backlog filename check. This package stays separate from existing courses and from the already-prepared article package.