Open the complete text-first lecture packageSeven-minute bridge: inverted index, document score, candidate recall. The current equivalent is notes, slide script, worked case, and no-video transcript; no recording is claimed.
Readiness check
Calculate Recall@5 from one supplied qrel and ranked list.
Bring
A one-page pipeline sketch with measurable and unknown denominators.
03
Explanation
Concepts, assumptions, and boundary
Eligibility is not exposure
Robots rules, sitemaps, HTTP responses, canonicalization, and indexing policies affect whether a source can enter a system’s available corpus. None guarantees that a particular query retrieves it. Retrieval then maps a query to candidates using sparse lexical matching, dense representations, or hybrid combinations. A useful audit records counts at each observable boundary—submitted URLs, successful fetches, indexed documents, retrieved candidates—and does not replace an unknown closed-platform denominator with a convenient proxy.
Baselines make failure visible
A sparse baseline such as BM25 is not included because it is always best. It provides an interpretable reference: term overlap, document length, and query wording can be inspected directly. Dense retrieval can improve semantic matching but introduces model, pooling, domain, and embedding-version dependencies. Hybrid retrieval can inherit strengths and failure modes from both. Topic-level error analysis matters more than a single mean because a system may gain on paraphrases while losing exact entities, dates, or rare terminology.
04
Primary visual
Inspect the mechanism or evidence structure
MechanismCandidate funnel with explicit loss points
Figure design. A four-stage funnel for crawl, index, retrieve, and context candidacy; each stage has numerator/denominator placeholders and a named “unknown” state. Each narrowing needs its own numerator, denominator, and unknown state.Long description
Documents narrow through four stages. Beside every narrowing is a count pair and a possible failure class. Unknown closed-system counts remain visibly blank rather than being estimated from the final response.
05
Interactive check
One action, one feedback state
Action
Compare frozen BM25, dense, and hybrid top-k run files by topic.
Feedback
Selecting a topic reveals false negatives, vocabulary mismatch, and whether the difference occurred before or after candidate generation.
Accessible alternative
All ranked lists and metrics are duplicated in a sortable text table.