Skip to content

FAQ (for stakeholders)

This page is the canonical reference for FAQ (for stakeholders).

Who this is for

  • PMs, engineering leaders, and procurement stakeholders evaluating RecSys
  • Anyone who needs “plain English” answers before diving into the technical docs

What this answers

  • What RecSys does (and does not) replace
  • What you need to run a pilot
  • How you control results and measure lift
  • How rollbacks and safety work operationally

Questions

Will this replace our search?

No. Search is intent-driven (“I want X”), while recommendations are typically discovery-driven (“you might like Y”). In most products, they work together:

  • Search solves “find what I asked for”.
  • Recommendations solve “help me decide / help me discover”.
How do we control what shows up?

The suite provides a control plane and policy hooks:

  • Rules: pin / boost / block by surface and segment.
  • Constraints: required/forbidden tags and per-tag caps.
  • Allow-lists / exclude lists: constrain candidate sets per request when needed.

For “why did we show this?”, use explainability options during development (options.include_reasons / options.explain).

How do we measure lift and make shipping decisions?

You measure recommendations with logs:

  • log exposures (what you showed)
  • log outcomes (click/conversion)
  • join by request_id

Then choose an evaluation mode:

  • Offline evaluation: fast regression gate before shipping.
  • A/B experiments: business KPI lift and guardrails for shipping decisions.
  • Interleaving / OPE: advanced options when experiments are slow or hard.
What do we need for a pilot?

Minimum viable pilot (DB-only mode):

  • a small catalog (items + tags)
  • a minimal popularity signal (daily rows)
  • exposure + outcome logging for evaluation

This is intentionally lightweight so teams can validate the end-to-end loop before adding pipelines and artifact mode.

How do rollbacks work?

Two rollback levers exist:

  • Config/rules rollback: versioned control-plane documents stored in Postgres.
  • Manifest rollback (artifact mode): “ship” and “rollback” are manifest pointer updates; artifacts are immutable.
What about privacy and compliance?

RecSys is designed to work with pseudonymous identifiers. The service can hash identifiers for exposure logging (do not log raw PII), and you can configure retention policies for logs.