Strategies generate candidates
Strategies build pools of candidates in the background — what's trending, what people engaged with together — with no per-person targeting. That work is shared across everyone.
Bosca Recommendations
Bosca Recommendations generates candidates with strategies you tune, then ranks them live for each viewer — blending behavior, meaning, and cohort. It works on brand-new content from day one, and learns from every gesture.
How it works
The work splits in two: building candidate pools is global and shared; ranking them is personal and live. That's what keeps personalization fresh without precomputing a feed for everyone.
Strategies build pools of candidates in the background — what's trending, what people engaged with together — with no per-person targeting. That work is shared across everyone.
When a person opens a feed, the platform ranks the candidates for that viewer on the spot — blending the model's scores, their ratings, and what they've hidden. Only people who ask for a feed pay for it.
A gesture — boost, lower, care, or hide — feeds straight back into the next ranking, so the feed keeps adjusting to what each person actually wants.
What makes it powerful
A production recommender — cold-start, multi-signal ranking, quality gates, and experimentation — wired into the same content, profiles, and events as the rest of the platform.
Feeds are ranked on demand, not precomputed for every profile in a batch. Ranking, rating-based re-ranking, and dismissal filtering all run at request time — so only the people who ask for a feed cost anything to serve.
New content is recommendable immediately. A model that learns from a piece's own features and meaning surfaces it before it has a single interaction.
Behavioral co-engagement, semantic similarity, and cohort personalization merge into a single ranked set — and every result can explain which signal put it there.
Boost, lower, care, and hide gestures become ratings and dismissals that steer the next ranking. High ratings even become durable learned interests.
Co-engagement can be conditioned on a viewer's cohort — a deterministic key built from the signals you define — so a feed aligns to people who behave alike.
A newly trained model has to measure up to the one in service on held-out data before it can take over; per-placement score floors and diversity caps keep low-confidence and repetitive results out.
Provision an experiment to pit the model against a heuristic, or a new model against the current one — split by a feature flag, with a safe default if anything is off.
Try every surface live from Studio against any profile or item — feeds, trending, similar, co-engagement — each result tagged with its score, strategy, and reason.
Serving surfaces
Every surface is a query on the same recommendation engine — a feed, a "related" row, a "people also viewed" strip — so you place the right one wherever it belongs.
forYouA person's personalized feed, ranked live from the trained model.
placementRecommendations for a named spot in your app, assembled from its strategies.
recommendedThe "related content" surface — behavioral and content signals merged, re-ranked for the viewer.
similarContent alike in meaning, from the content model's item-to-item index.
coEngaged"People who engaged with this also engaged with" — item-to-item behavior.
trendingGlobally trending content by recent interaction velocity.
Go deeper
The candidate generators — trending, co-engagement, and the live model.
Named surfaces that assemble a ranked set for any spot in your app.
Turn profile attributes and segments into what personalizes a feed.
The trained recommender, its quality gates, and built-in A/B testing.
The recommender learns from the same content, profiles, and events every other subsystem produces — no pipeline to build, no data to ship out. The docs cover strategies, placements, signals, and the model end to end.