Reggie Preference Learning

Automatic learning of user preferences and repetitive workflows to make Reggie smarter over time.

Overview

The Preference Learning System runs a lightweight analyzer in parallel with every Reggie chat message. It detects:

Preferences - How the user wants responses formatted/styled
Workflows - Repetitive tasks the user does regularly
Pain Points - When users re-explain or express frustration
Corrections - When users explicitly correct Reggie's behavior

These are stored and automatically injected into future conversations, making Reggie progressively smarter about each user.

Architecture

┌─────────────────────────────────────────────────────────┐
│                    User Message                         │
└─────────────────────┬───────────────────────────────────┘
                      │
         ┌────────────┴────────────┐
         │                         │
         ▼                         ▼
┌─────────────────┐    ┌─────────────────────────┐
│  Main LLM Call  │    │  Preference Analyzer    │
│  (GPT-5/Pro)    │    │  (GPT-4.1-nano)         │
│                 │    │  - Detects preferences  │
│                 │    │  - Matches workflows    │
│                 │    │  - Logs patterns        │
└────────┬────────┘    └───────────┬─────────────┘
         │                         │
         │                         ▼
         │             ┌─────────────────────────┐
         │             │  Store to Database      │
         │             │  - user_llm_preferences │
         │             │  - user_learned_workflows│
         │             │  - user_prompt_patterns │
         │             └─────────────────────────┘
         │
         ▼
┌─────────────────────────────────────────────────────────┐
│  Response (with preferences/workflow context injected)  │
└─────────────────────────────────────────────────────────┘

What Gets Learned

Preferences (Response Style)

Category	Examples
`tone`	"Be more casual", "No emojis", "Professional tone"
`length`	"Be concise", "Give detailed explanations", "Bullet points only"
`format`	"Always show code examples", "Include SQL queries", "Step-by-step"
`expertise`	"I'm a senior dev", "Skip the basics", "Explain like I'm new"
`output`	"Always show your reasoning", "Just give me the answer"
`behavior`	"Don't apologize", "Ask clarifying questions first"

Workflows (Repetitive Tasks)

When the system detects a user asking similar questions 3+ times, it creates a learned workflow:

{
  "workflow_name": "roster_vehicle_extraction",
  "trigger_pattern": "roster CSS/file + vehicles",
  "learned_steps": [
    "Parse CSS table structure",
    "Find vehicle column", 
    "Extract assignments by row",
    "Format as summary list"
  ],
  "preferred_output": "bullet list with times"
}

Next time the user drops a roster file, Reggie proactively knows what to do.

Data Model

`admin.user_llm_preferences`

- id UUID
- user_id UUID (FK)
- category TEXT (tone, format, length, expertise, output, behavior)
- preference TEXT (short identifier)
- description TEXT (human-readable for UI)
- source_message TEXT (what triggered this)
- confidence FLOAT (0-1)
- times_reinforced INT (increases when re-detected)
- active BOOLEAN
- created_at, updated_at

`admin.user_learned_workflows`

- id UUID
- user_id UUID (FK)
- workflow_name TEXT
- trigger_pattern TEXT
- learned_steps JSONB
- preferred_output TEXT
- example_input TEXT
- example_output TEXT
- times_detected INT
- last_used_at TIMESTAMPTZ
- confidence FLOAT
- active BOOLEAN
- created_at, updated_at

`admin.user_prompt_patterns`

- id UUID
- user_id UUID (FK)
- session_id UUID (FK)
- prompt_hash TEXT (MD5 for similarity matching)
- prompt_summary TEXT
- prompt_category TEXT
- created_at

Rolling window of last 100 prompts per user for pattern detection.

Context Injection

When a user sends a message, their preferences and any matching workflow are injected:

[User Preferences]
• Prefers concise responses
• Senior developer - skip basic explanations
• Always include code examples

[Known Workflow: roster_vehicle_extraction]
This user regularly does this task. Their preferred approach:
- Trigger: roster CSS/file + vehicles
- Output format: bullet list with times
You can proactively help without them re-explaining.

API Endpoints

GET `/api/pdb/v1/reggie/preferences`

List all learned preferences for current user.

GET `/api/pdb/v1/reggie/workflows`

List all learned workflows for current user.

GET `/api/pdb/v1/reggie/context`

Get combined preferences + workflows (preview what gets injected).

PATCH `/api/pdb/v1/reggie/preferences/:id/toggle`

Toggle preference active state.

{ "active": false }

DELETE `/api/pdb/v1/reggie/preferences/:id`

Remove a learned preference.

PATCH `/api/pdb/v1/reggie/workflows/:id/toggle`

Toggle workflow active state.

DELETE `/api/pdb/v1/reggie/workflows/:id`

Remove a learned workflow.

POST `/api/pdb/v1/reggie/preferences`

Manually add a custom preference.

{
  "category": "output",
  "preference": "include_sql",
  "description": "Always show SQL queries"
}

UI Integration

A cog icon in the Reggie chat header opens a preferences modal:

┌─────────────────────────────────────────────┐
│  Your Reggie Preferences            [x]     │
├─────────────────────────────────────────────┤
│  Learned from your conversations            │
│                                             │
│  ☑ Concise responses preferred              │
│  ☑ Senior developer - skip basics           │
│  ☑ Always show SQL queries                  │
│  ☐ Include code comments         [Remove]   │
│                                             │
│  ─────────────────────────────────────────  │
│  Learned Workflows                          │
│                                             │
│  ☑ Roster Vehicle Extraction (used 12x)    │
│  ☐ Staff Scheduling Report (used 3x)       │
│                                             │
│  + Add custom preference                    │
│           [Save Changes]                    │
└─────────────────────────────────────────────┘

Analyzer Prompt

The GPT-4.1-nano analyzer receives:

{
  "current_message": "user's message",
  "similar_past_prompts": [
    { "summary": "...", "days_ago": 2, "similarity": 0.85 },
    ...
  ],
  "match_count": 4,
  "recent_context": [ /* last 4 messages */ ]
}

And returns:

{
  "detected": "preference|workflow|pain_point|correction|none",
  "preference": {
    "category": "tone",
    "preference": "concise",
    "description": "Prefers concise responses",
    "confidence": 0.85
  },
  "reasoning": "User said 'just give me the short version'"
}

Configuration

Env Variable	Default	Description
`REGGIE_PREFERENCE_LEARNING`	`true`	Enable/disable the system

The analyzer uses gpt-4.1-nano for speed and cost efficiency (~$0.20/1M tokens).

Integration with Context System

This system extends the existing Context v2.1:

v2.0: Session memory, recent/history summaries, topic inference
v2.1: LLM summaries, semantic search, personalized intros
v2.2: Preference learning, workflow detection, implicit task memory

All three layers work together:

[Personalized Intro - v2.1]
"Hey Lloyd! Last time we worked on the roster analysis..."

[Recent Context - v2.0]
• Previous session summaries...

[User Preferences - v2.2]
• Concise responses preferred
• Senior developer

[Known Workflow - v2.2]
• roster_vehicle_extraction (matched)

[Topic: Scheduling - v2.0]
• Topic-specific context...

Future Enhancements

Periodic check-ins: "Quick check - are these preferences still accurate?"
Workflow sharing: Share common workflows across team
Confidence decay: Reduce confidence over time if not reinforced
Category detection: Auto-categorize prompts (analysis, generation, lookup)
Cross-session learning: Learn from corrections in real-time

Overview​

Architecture​

What Gets Learned​

Preferences (Response Style)​

Workflows (Repetitive Tasks)​

Data Model​

admin.user_llm_preferences​

admin.user_learned_workflows​

admin.user_prompt_patterns​

Context Injection​

API Endpoints​

GET /api/pdb/v1/reggie/preferences​

GET /api/pdb/v1/reggie/workflows​

GET /api/pdb/v1/reggie/context​

PATCH /api/pdb/v1/reggie/preferences/:id/toggle​

DELETE /api/pdb/v1/reggie/preferences/:id​

PATCH /api/pdb/v1/reggie/workflows/:id/toggle​

DELETE /api/pdb/v1/reggie/workflows/:id​

POST /api/pdb/v1/reggie/preferences​

UI Integration​

Analyzer Prompt​

Configuration​

Integration with Context System​

Future Enhancements​