Table of Contents
🌏 中文版
Units 13–14 move from model internals to system boundaries: knowledge sources, segmentation and indexing, evidence delivery to a generator, and architectural trade-offs. The useful reading strategy is not memorizing a RAG diagram, but following Assignment 3's failure decomposition.
Four separable RAG decisions
Separate corpus, retrieval unit, retriever, and generator. The corpus sets scope and timestamp. Chunking sets evidence granularity. Sparse, dense, or hybrid retrieval forms candidates. Only then does the generator turn question and evidence into an answer. Changing all four at once destroys attribution.
A3 intentionally has no starter code
Assignment 3 asks teams to answer short questions about Berkeley EECS pages. They crawl and clean the site, construct validation data, check annotation agreement, build a corpus, and use Exact Match, token F1, and retrieval recall for ablations.
The grading environment has no GPU, 4GB RAM, fixed dependencies and entry points, model-size limits, and latency constraints. The largest model is therefore not necessarily the deliverable system. A self-study version can use a smaller question set, but should retain timestamps, source URLs, retrieval recall, and end-to-end metrics.
Access and safety boundaries
The PDF is public; hidden data, the official wrapper, and Gradescope are not. Solutions must not be distributed. Impact and Social Implications appears on the schedule without anonymous slides, so this article recommends recording crawl scope, personal-data handling, refresh policy, and citations without pretending those recommendations summarize the private lecture.
Try this tonight: crawl ten public EECS pages, write ten short-answer questions, build a BM25 baseline, save top-k evidence, and classify each failure as retrieval or generation.
References
Loading...