Learning Rules

​

What learning rules do

When you connect integrations like HubSpot, Zoom, Fathom, or PandaDoc, Halo can ingest a lot of data — every email engagement, every meeting transcript, every signed contract. Some of it is gold for your AI. Some of it is noise. Learning rules let you tell Halo what topics to extract knowledge from and what to skip — without changing the source of truth in those tools. Configure them at Settings > AI Agents > Knowledge Learning Rules.

​

Three rule types

​

Topic rules

Topic rules use the AI to filter incoming content. When a HubSpot engagement, Zoom transcript, or other distillable source comes in, Halo asks: “Does this match the topic?” If yes, the content is distilled and added as knowledge. If no, it’s skipped. Use topic rules when you want the AI to learn only from specific kinds of content:

• “Customer pain points and feature requests”
• “Product feedback and bug reports”
• “Pricing discussions”
• “Onboarding conversations”

You can have multiple topic rules. Content matching any rule is ingested.

​

Keyword rules

Keyword rules are exact-match filters that prevent content from being learned. If the keyword (or phrase) appears anywhere in the source, Halo skips it. Use keyword rules when you have well-defined exclusions:

• confidential — skip anything marked confidential
• internal only — skip internal-only meetings
• personal — skip personal notes

Keyword rules are case-insensitive and match anywhere in the content.

​

Instruction rules

Instruction rules are extra extraction guidance that gets passed to the AI when distilling synced content. They don’t filter — they shape how distillation works. Examples:

• “When ingesting Fathom transcripts, focus on action items, customer commitments, and product feedback. Ignore small talk.”
• “For HubSpot engagements, extract any pricing discussions verbatim.”
• “When learning from PandaDoc contracts, capture renewal dates, contract terms, and special clauses.”

Instruction rules give you fine control over what the AI distills without writing custom integrations.

​

Rule limits

There’s a workspace-wide cap on the total number of learning rules and the length of each rule’s value. Halo will warn you in the UI when you’re near the limit. If you find yourself wanting to add many specific exclusions, consider using broader topic or instruction rules instead.

​

When rules apply

Learning rules run on distillable sources — currently:

• HubSpot engagements (emails, notes, calls, meetings)
• Zoom meeting transcripts
• Fathom meeting transcripts
• PandaDoc contracts
• Intercom imported conversations

They do not affect:

• URL crawls (use crawler config or robots.txt instead)
• File uploads (manual control)
• Internal docs (manual control)
• Slack channel ingest (separate per-channel mapping)

​

Tuning over time

Like instructions and escalation, learning rules benefit from iteration:

1. Start with one or two topic rules that describe what you want the AI to focus on.
2. Add keyword rules as you discover content the AI is wrongly learning from.
3. Add instructions when you want richer or more focused distillation.

Watch your Knowledge tab to see what’s being indexed. If too much noise is showing up, tighten the rules. If important context is being skipped, loosen them.

​

Where to go next