←
AI/ML Systems
Deriva
Learn
AI/ML
Patterns
Observe
Search
⌘K
More
⌄
0%
AI/ML Systems
0%
Home
Learn
Patterns
Observe
More
Opening…
AI/ML
›
Track
SEARCH-001 – REC-005
Retrieval, ranking, and recommendation
0 of 20 done · 0 due for review.
all
new
started
attempted
done
all kinds
implement
derive
debug
construct
predict
compare
communicate
SEARCH-001
Build an inverted index for these five documents: map each term to the documents containing it.
implement · search-before-generation
new
SEARCH-002
Calculate term frequency and inverse document frequency for 'crash' in 3 of 10 documents.
derive · search-before-generation
new
SEARCH-003
Rank these documents with BM25-style evidence for the query 'crash fix': score = sum of idf × tf-weight over matching terms.
implement · search-before-generation
new
SEARCH-004
Diagnose this lexical retrieval miss: the query 'dashboard shows zero' returns nothing, but a ticket about 'revenue display bug' should match.
debug · search-before-generation
new
SEARCH-005
Design a retrieval evaluation set: 20 queries, relevant documents per query, and the metric to report.
construct · search-before-generation
new
VECTOR-001
Calculate nearest-neighbor similarity: doc A [1, 0, 0], doc B [0.8, 0.2, 0], query [1, 0, 0.1]. Which doc is closer, by cosine?
derive · similarity-is-a-representation-choice
new
VECTOR-002
Diagnose this embedding-space collision: 'refund policy' and 'refund denied' map to nearly identical vectors and rank together.
debug · similarity-is-a-representation-choice
new
VECTOR-003
Choose an index for latency versus freshness: 10M vectors, updates every minute, p99 under 50 ms.
predict · similarity-is-a-representation-choice
new
VECTOR-004
Detect stale vectors after these document updates: 40 docs were edited, but search still returns the old versions.
debug · similarity-is-a-representation-choice
new
VECTOR-005
Compare lexical and semantic retrieval for this query: 'app crashed' vs 'application terminated unexpectedly'.
compare · similarity-is-a-representation-choice
new
RANK-001
Separate candidate recall from final ranking: 50 candidates come from retrieval, 10 must reach the user. Design the two stages.
construct · ranking-is-a-separate-responsibility
new
RANK-002
Choose ranking features for this query: a user searching 'bug' in tickets. Pick three features that should order the results.
construct · ranking-is-a-separate-responsibility
new
RANK-003
Debug the reranker that reverses strong evidence: the top candidate has 3 exact keyword matches but ranks 8th.
debug · ranking-is-a-separate-responsibility
new
RANK-004
Calculate this ranking metric: relevant docs at positions 1, 3, 5 for a query. Compute MRR.
derive · ranking-is-a-separate-responsibility
new
RANK-005
Explain this relevance-label disagreement: two annotators rated the same (query, doc) pair 3 and 1 out of 5. Write the resolution.
communicate · ranking-is-a-separate-responsibility
new
REC-001
Build a popularity baseline: rank tickets by total views, ties broken by recency. Write the ordering rule.
implement · ranking-is-a-separate-responsibility
new
REC-002
Diagnose the cold-start problem: a brand-new ticket gets no views, so popularity ranks it last, so nobody sees it.
debug · ranking-is-a-separate-responsibility
new
REC-003
Identify feedback-loop amplification: the recommender promoted ticket X, views grew, so the recommender promoted X more.
debug · ranking-is-a-separate-responsibility
new
REC-004
Choose exploration versus exploitation: 90% of traffic ranks by the learned model, 10% by recency/random. Predict what the 10% buys you.
predict · ranking-is-a-separate-responsibility
new
REC-005
Design a recommendation slice evaluation: mobile vs desktop users, new vs returning.
construct · ranking-is-a-separate-responsibility
new