Pipeline

Every recommendation request runs through a pipeline: an ordered set of stages that retrieve candidates, score them, apply your rules and shape the final list. Each context has its own pipeline, created with its recipe, so a homepage feed and a “similar items” rail can work differently. Open a pipeline from a context’s recipe, or from Build a pipeline on the Dashboard. The editor shows each stage as a card that you can turn on or off and configure. When opened from a context, the editor is locked to that context’s pipeline.

​

Stages

Requests pass through the stages in this order. Turning a stage off skips it entirely.

​

Stage settings

​

Candidate generation

• Candidate limit: how many items to retrieve before scoring. Higher values widen the candidate pool and can increase latency. Evaluate a larger pool if results lack variety.
• Query / search retrieval: for search requests, adds keyword matching alongside meaning-based matching. Choose Reciprocal rank fusion (rank-based and stable) or Weighted normalised scores, set the semantic and keyword weights, and pick the keyword fields (built-in fields and metadata keys) to match against.

​

Scoring

​

Learned ranking

Choose the ranker type: Ranking facets tells the ranker which metadata fields to compute topic affinity over, such as category or brand. The first non-empty field on each item is used. Leave it empty to use standard category and genre keys.

​

Catalogue intelligence

Blends relationships between items (similar, complementary, substitute, same collection, frequently used together, next best and others) into model scores.

• Catalogue influence: how far these signals can shift scores, capped at 50% so model relevance stays primary.
• Relationship types and sources: which relationships may be used.
• Max relationship candidates: how many relationship-derived candidates may enter per request.

​

Intelligence layer

An optional reranker for the top of the list.

​

Rules

Turn this stage off to serve pure model ranking with no business rules, for example as the control arm of an experiment.

​

Post-processing

• Diversity: limits consecutive items from the same category, pushing extras further down.
• Deduplicate: removes duplicate items, keeping the highest-scored one.
• Page size: how many items the response returns. Every earlier stage works on the full candidate set; the trim happens here.

​

Video understanding

If Twelve Labs video enrichment is enabled for your workspace, the pipeline can use the stored video analysis for candidate vectors, with Fallback to base embeddings keeping recall stable for videos not yet enriched.

​

Defaults

A new context’s pipeline comes from its template. A workspace with no pipeline at all uses defaults: all stages on, a candidate limit of 100 and a page size of 20.

​

Pipelines and experiments

An experiment variant can use a different recipe, and with it a different pipeline, so you can measure the effect of a stage. For example, run the rules stage off in one variant to see whether your rules help. See Experiments.