Pipeline Config

The pipeline config page gives you control over the multi-stage recommendation pipeline. Each recommendation request flows through a series of stages, and you can enable, disable, or tune each one independently.

---

​

Pipeline stages

Recommendations are produced through these stages, in order:

---

​

Configuring the pipeline

​

When to disable stages

• Disable Rules to produce pure ML-ranked results. Useful as a control arm in an A/B experiment to measure whether your business rules help or hurt engagement.
• Disable Feature Enrichment if you are testing raw embedding similarity without metadata features.

​

Tuning candidate limits

The candidate_limit in the candidate generation stage controls how many items are pulled from the vector index before scoring. A higher limit gives the scoring model more candidates to choose from but increases latency. Start with 100-200 and increase if you notice the final results lack diversity.

​

Tuning page size

The page_size in post-processing determines how many items are returned in the API response. This does not affect internal processing — all stages operate on the full candidate set. The trim happens at the end.

---

​

Pipeline and experiments

When running an A/B experiment, each variant can reference a different pipeline config. For example:

• Control: default pipeline with all stages enabled
• Treatment: pipeline with the rules stage disabled

This lets you isolate the impact of specific pipeline stages on recommendation quality. See A/B Testing for setup details.

---

​

Fallback behaviour

If no pipeline config exists for your team, the system uses sensible defaults:

• All stages enabled
• candidate_limit: 100
• page_size: 20

You only need to configure the pipeline if you want to change these defaults.