Skip to content

QiLife AI Review Workflow

To prevent AI hallucinations, model drift, or unintended edits from corrupting Cody's personal ledger, QiLife operates on a strict Human-in-the-Loop (HITL) doctrine. The AI proposes; Cody approves or edits.

1. Core Principles

  1. No Silent Creation: The AI service never directly inserts or updates records in primary tables (actions, obligations, transactions, people, threads) without user intervention.
  2. Draft / AI Output Staging: All AI inferences are written to the ai_outputs table first. They are displayed in the UI as editable suggestions.
  3. Explicit Acceptance: A suggestion only moves to the primary tables when Cody clicks Accept (which marks ai_outputs.accepted = true and performs the writes) or modifies the values and clicks Save.
  4. Provenance Tracking: Created items retain a link to their originating qibit_id and ai_output_id via the links table.

2. The Triage Review Pipeline

The primary review pipeline happens in the Inbox during the Triage phase of a QiBit.

  [Raw Capture]
        ↓
   AI Service (Calls interpret_qibit & extract_related_entities)
        ↓
  [Insert into ai_outputs] (State: accepted=false)
        ↓
  [Inbox UI Display] (Show suggested fields side-by-side with original text)
        ├─────────────────────────────┐
        ▼                             ▼
  [Cody clicks Edit]            [Cody clicks Accept]
        ▼                             ▼
   Modifies fields in form      Backend writes direct to primary tables
        └──────────────┬──────────────┘
                       ▼
            [Commit to SQLite]
     - Mark ai_outputs.accepted = true
     - Create records in primary tables
     - Create links for provenance

3. Database State Machine

The ai_outputs Table State

  • accepted = false (default): Represents a pending suggestion. It is displayed as a draft card in the inbox or context dock.
  • accepted = true: Cody has approved this output. The application logic prevents writing this output to the primary tables again to avoid duplicate creation.

When a suggestion is accepted: * A link is written: source_type = 'qibits', source_id = [QiBit ID], target_type = 'actions', target_id = [Created Action ID], relationship = 'created_from'. * If money is extracted: source_type = 'qibits', source_id = [QiBit ID], target_type = 'obligations', target_id = [Created Obligation ID], relationship = 'created_from'.

4. UI Representation of AI State

The frontend must visually distinguish between three categories of data: 1. Confirmed Data (User Entered/Approved): Standard color coding, fully indexed. 2. Suggested Data (AI Drafts): * Displayed with dashed borders or light purple background. * Accompanied by a confidence rating indicator (e.g., Confidence: 85%). * Features a quick-action checkmark (Accept) and a pencil icon (Edit). 3. Contradictions / Flags: If an AI background process detects an unresolved open loop or a duplicate entry, it places a red warning icon in the Context Dock under "AI Flags".