Recommender & Ranking Archetype

A typed query is a retrieval constraint applied before any model runs, so the ranker only sees items that already match the stated intent. A browse feed has no such constraint and must manufacture candidates from the viewer's profile and context.

A signal test on the request, not the account's age. Each rung declares a precondition; the funnel serves the first rung whose precondition holds, re-checked on every request, and logs which rung answered.

Cluster the near-duplicates first, then select greedily. Content similarity over title and lead text groups the forty filings into one event cluster; the list-editing pass then picks one representative and, using maximal marginal relevance, scores each further slot on relevance minus similarity to what is already placed.

Inside retrieval, the shortlist comes back full of eligible items. Filtering afterwards spends the shortlist on ineligible ones, so a selective facet leaves too few results to fill the page and wastes the scoring stage's budget.

Features that mix user, clip and request context in one value - this viewer's past finishes of this creator, the slot, the time of day, the connection. Retrieval must fix a clip's representation before the request arrives, so it cannot hold any of them.

Measure two different recalls: the index's recall against exact vector search, and the funnel's recall against the scorer's own top-25 over a much deeper candidate pool replayed offline. Clicks cannot measure it — a posting never retrieved was never shown.

By reserving impressions rather than hoping for them: a fixed share of slate positions goes to items below a confidence bar, metered by a per-item impression quota so the budget spreads across inventory, and each item graduates once its estimate is trustworthy.

At the last stage before the response: a suppression check against the final slate, reading a takedown list from a low-latency key-value store on the serving path. Removing the article upstream is also right, but candidate sources and precomputed slates refresh too slowly to be the guarantee.

Because a blend adds magnitudes across heads. Each head can order its own predictions perfectly while its numbers are systematically inflated, and the inflated head then dominates the sum - so the weighted order, any fixed cutoff and any value arithmetic come out wrong.

A per-outlet slot quota in the list-editing pass that runs after scoring. The ranker orders candidates by relevance score; the quota then walks that order and skips any outlet that already holds two slots on the page.

The retrieval score decides membership - whether a clip enters the shortlist at all. The ranking score decides order - where a shortlisted clip sits in the slate. The second stage may reverse retrieval's order entirely.

Merge on rank, not on raw score: cosine similarity, recency order and an alert match are incomparable numbers. Reciprocal rank fusion or fixed per-source quotas produce one list; collapse duplicates on a canonical posting id before truncating.

A representation built from what exists at publish time — declared category, description text, host and network links, transcript topics. It is either projected into the retrieval space alongside behavioural items or run as a separate candidate source with its own quota in the merge.

The cap's window is the page, and the window resets at the page boundary. Slots nine and ten of the first page plus slots one and two of the second are four in a row to a scrolling reader, because the second request counted from zero.

Split additions from removals. New postings are embedded on arrival into a small fresh segment searched beside the nightly-built base index; filled or expired postings are dropped on the read path from a tombstone set, because the structure only reclaims them at rebuild.

Two partitions for two different reasons. Tenant is a correctness boundary: a private board gets its own index selected before the search, never a filter applied afterwards. Locale is a performance partition inside the public index, routed by the seeker's commute area.

Only until observed behaviour replaces it. Treat the picks as a low-confidence prior that seeds retrieval while there is nothing else, then decay their weight against completed listens — not against elapsed days — and keep them attributable rather than written in as fake events.

The two scores share no scale. A keyword score has no fixed ceiling and moves with query length and term rarity, while a similarity score is bounded but query-dependent, so per-query normalisation makes a weak list look exactly like a strong one. Rank fusion merges on position instead.

A typed query asserts an intent the catalogue may not satisfy, and retrieval will hand back its nearest candidates regardless, so without a cutoff the surface silently answers a different question. A browse feed makes no such assertion to contradict.

It never contains outcomes for clips retrieval missed or the ranker buried, because nobody saw them. Those are unobserved, not negative. The model ends up confident where the funnel already agreed and evidence-free exactly where it must extrapolate.

Subtract the stage's fixed cost from its budget and divide by the per-candidate cost. With 6 ms of batched feature fetch and setup, 24 ms remain, so about 160 clips fit. Shortlist size is arithmetic, not taste.

Make precedence a property of the system, not of insertion order; give every rule an owner, a reason and an expiry; measure each rule's cost with holdouts and per-rule attribution logging; and ship and retire a rule through the same ramp and flag a model change gets.

Retrieval silently returns near-random postings. Each training run produces its own coordinate space, so a new seeker vector compared against the previous encoder's posting vectors is a meaningless comparison — with no error, unchanged latency and a full result set.

Because the two change on different timescales. A post-scoring multiplier is an operations lever an editor can turn during a breaking event and reverse in minutes; article age as a model input learns a better-shaped response from data but only moves when the model is retrained.

The two surfaces label different things. A feed click signals interest; a search click signals a match to a stated intent, and an abandoned query is a failure on search but an ordinary session on the feed. Pooled, the higher-traffic surface's objective wins.

Evidence that the current reserve no longer bootstraps the inventory arriving: a graduation rate below the publishing rate, p95 time-to-first-impression drifting out, falling catalogue coverage, or supply-side churn among creators who never get shown.