Improvement Pipeline

The Knowledge Dialog includes a Review tab with an automated improvement pipeline for maintaining knowledge quality over time. The pipeline runs in four stages — Audit, Enrich, Deduplicate, and Refine — each designed to identify and fix different classes of issues.

Audit

The audit stage scans all knowledge items and flags potential issues. It produces a report organized by category so you can see the overall health of your knowledge base at a glance.

• Missing keywords — items that lack keyword entries, reducing their findability in search
• Stale or missing embeddings — items whose vector embeddings are outdated or have never been generated
• Duplicate clusters — groups of items that appear to describe the same thing
• Quality flags — items marked with low effectiveness, high failure rates, or other quality concerns
• Unused items — items that have never been recalled or applied in any conversation

Enrich

The enrich stage addresses items flagged during the audit by filling in missing data and regenerating stale content. You choose which flagged items to enrich, so you stay in control of every change.

• Add missing keywords — generates relevant keywords for items that lack them, improving retrieval accuracy
• Rebuild embeddings — regenerates vector embeddings for items with stale or missing data, ensuring up-to-date similarity search results
• User selection — you review the flagged items and select which ones to enrich before any changes are made

Deduplicate

The deduplicate stage identifies duplicate clusters — groups of knowledge items that describe the same concept — and generates merge proposals to consolidate them.

• Cluster detection — groups items that are semantically similar, even if they use different wording
• Merge proposals — for each cluster, an LLM-powered proposal shows:
  • Which item is the primary (kept) and which is secondary (merged into the primary)
  • A similarity score indicating how closely the items match
  • An LLM confidence rating for the proposed merge
  • A merge preview listing what would be added: keywords, triggers, indicators, or aliases
• Approve or reject — you review each merge proposal individually and decide whether to apply it

Refine

The refine stage runs a quality checklist against selected knowledge items, identifying specific issues and suggesting concrete fixes.

• Quality checklist — each item is evaluated against a set of criteria, with pass/fail results shown for every check
• Suggested fixes — when issues are found, each fix includes:
  • The field that needs attention
  • The current value
  • The suggested value
  • The reason for the change
• User selection — you choose which fixes to apply, keeping full control over what gets changed