Skip to content

Data Product Recommender

The Data Product Recommender is an unattended agent that runs the same evidence pipeline as the Data Product Generator — identify, reconfirm, synthesize, prioritize — but never creates a product. Instead, it writes recommendation cards to the Alation Opportunities Inbox so data stewards can review and claim them at their own pace.

This agent is designed to run without a human in the loop. It does not stop to ask for approval; it runs the full pipeline and writes the results.

  1. Reads the current Opportunities Inbox to avoid re-proposing already-active clusters
  2. Identifies candidate table clusters from relational database query usage and catalog metadata
  3. Reconfirms each cluster against BI usage data and per-table usage detail
  4. Groups tables into coherent clusters and ranks them by impact
  5. Writes all candidates to the Opportunities Inbox in a single batch

Required:

  • message (string): Describes the scope — “scan all active datasources” or “recommend data products for the SALES schema”

Optional:

  • datasource_id (integer): Scope the evidence pipeline to a single datasource

The agent produces thinking, tool call, tool return, and text blocks as it runs. It does not pause for input. The final message reports the outcome: how many cards were inserted, how many already existed, and how many were skipped.

Each card written to the Opportunities Inbox contains:

FieldDescription
titleBusiness-outcome name (e.g. “Sales Orders”), not a schema or table name
short_descriptionOne line leading with business value and the BI reports it supports
sourceThe primary signal: query_history (default) | bi_usage | strategic_insights
confidencehigh | medium | low based on usage strength and BI coverage
member_tablesNon-empty list of fully-qualified table names
evidenceUsage context, BI consumers, confidence rationale, and suggested example questions
  • Checks the inbox first — skips any candidate that substantially overlaps an existing active card
  • Writes all candidates in a single batch; skipped candidates are not retried
  • Individual query details and user identities from usage data are never exposed
  • Gaps in BI usage data are noted in the evidence rather than being treated as confirmation that no consumers exist