Intelligence / MLE

Overview

MLE (Beta) is a conversational AI assistant that lives in the Intelligence section of the agent dashboard. It knows the agent's listings, brand voice, service areas, and business context — agents never need to explain their business to it.

MLE helps agents with:

• Content generation — Social posts, listing descriptions, email drafts with iterative refinement
• Listing search — Natural language queries against MLS data ("Show me waterfront condos under $3M")
• Market analysis — Portfolio performance, inventory summaries, days on market
• Client preparation — Briefings and follow-up recommendations (Phase 2)

Dashboard Location

Intelligence
  ├── MLE (Beta)     ← Conversational assistant
  ├── Voice          ← Brand voice settings
  ├── Automations    ← AI content triggers
  └── Usage          ← Stats & limits

The Intelligence section consolidates all AI features that were previously spread across Marketing and AI Studio.

Conversation Types

Tone Selector

MLE responses can be tuned with four tone presets:

• Professional (default) — Polished and expert
• Casual — Warm and conversational
• Luxurious — Sophisticated, aspirational language
• Urgent — Concise and action-oriented

Feature Gating

Feature Flag

• features.intelligence: boolean — Enable/disable per workspace
• Set via admin panel: Workspace Settings > Intelligence / MLE

Architecture

System Prompt

MLE's system prompt is assembled dynamically per request:

1. Base persona — Luxury real estate assistant behavior rules
2. Fair Housing compliance — Mandatory prohibited language rules
3. Brand voice — Agent's tone, avoidances, writing sample
4. Workspace context — Brokerage, service areas, agent name
5. Listing data — Active/pending listings injected based on intent
6. Additional instructions — Platform-wide custom instructions (admin setting)

Intent Classification

User messages are classified to determine what context to pre-fetch:

• content_generation — Write/draft/create content
• listing_search — Find/show/search listings
• listing_detail — Tell me about a specific listing
• market_analysis — How is my inventory performing
• client_query — Ask about a client (Phase 2)
• general — Everything else

Streaming

Uses the Vercel AI SDK with streamText() and useChat() for real-time streaming responses.

Key Files

Database Tables

assistant_conversations (workspace_id, contact_id, title, type, pinned)
assistant_messages (conversation_id, role, content, metadata)

Both tables have RLS enabled (service-role-only pattern).

Admin Configuration

Global Settings

Location: Admin > Global Settings > Intelligence / MLE

• Enabled — Master kill switch (independent from AI Content Generation)
• Additional Instructions — Appended to MLE's system prompt for all workspaces
• Retention Days — Conversation cleanup policy (default 90 days)

Per-Workspace

Location: Admin > Workspaces > [workspace] > Settings > Intelligence / MLE

• Enabled — Toggle MLE for this workspace
• Usage stats — Conversations this month, remaining quota

Compliance

• Fair Housing Act rules are always enforced in the system prompt
• MLE operates only on behavioral data (no demographic reasoning)
• All responses include "based on MLS data" attribution for listing facts
• ai_usage table tracks token consumption per conversation
• assistant_messages.metadata stores model, intent, and token counts

Limitations (Beta)

• No client data integration yet (viewing/favoriting history) — planned for Phase 2
• No proactive intelligence (push notifications for matches) — planned for Phase 3
• No tool calling / platform actions (create note, save search) — planned for Phase 2
• Conversation history is not searchable yet
• No floating "Ask MLE" button on other dashboard pages — planned for Phase 2