​
Loop Simulator — Data Sources

The Loop Simulator at ella.navgurukul.org/v2/conversation-flow/ is a developer tool for testing the full learning loop. It has several panels on the right side. Here is where each panel gets its data.

​
Panel Data Sources

PanelSourcePersisted?
MemoryGET /users/me/memory (backend DB)Yes
API CallsBrowser log of all fetch() calls to the backendNo — lost on refresh
SessionAssembled locally from judge response + chat messagesNo — lost on refresh
Due_nextFrom GET /users/me/memory responseYes (in DB)
SkillsFrom curriculum_data.py static data (rendered client-side)Static — never changes
HistoryAccumulated locally across multiple chats in one sessionNo — lost on refresh

​
What Is Persisted vs Ephemeral

Persisted in PostgreSQL (survives refresh, device switch, etc.):
  • User profile (name, age, city, level)
  • skills_progress (owned, emerging, sightings, strength)
  • due_next (what the next session targets)
  • persona_memory (facts about the student)
  • Messages (every conversation turn)
  • Session analyses (judge scores, state changes)
Ephemeral (browser memory only — lost on page refresh):
  • Session logs in the Session tab
  • API call logs in the API Calls tab
  • History chart data
  • Chat messages displayed on screen (until saved to DB by turn endpoint)

​
The Learning Loop (API sequence)

1. POST /session         → Start session (auto-picks topic from due_next)
2. POST /session/{id}/turn  → Conversation turns (repeated 5-6 times)
3. POST /llm/judge       → End session + grade skills + update progress
   └─ Updates: skills_progress, due_next, user.level (in DB)
4. GET /users/me/memory  → Read updated state for next session
5. Back to step 1

​
Session Tab Structure

Each entry in the Session tab is assembled locally after the judge grades a conversation: 
{
  "session_id": "s_0001",
  "child_id": "device_uuid",
  "level": "A1",
  "progression_model": "unit",
  "prompt_version": "backend_v1",
  "timestamp": "2026-06-15T14:41:56Z",
  "turns_total": 5,
  "topic": "My name and where I'm from",
  "aimed_at": ["U1-VOC-01", "U1-GRA-01"],
  "hit_targets": ["U1-VOC-01"],
  "skills_evidenced": [
    {
      "id": "U1-VOC-01",
      "name": "I can say my name, age, and where I'm from.",
      "confidence": 0.8,
      "prompted": false,
      "evidence": "I am Ravi. I from India."
    }
  ],
  "state_changes": [
    {
      "id": "U1-VOC-01",
      "from": "emerging",
      "to": "owned"
    }
  ],
  "transcript": [
    {"speaker": "zoe", "text": "Hi! What's your name?"},
    {"speaker": "learner", "text": "I am Ravi."}
  ]
}
This data comes from the /llm/judge response combined with the local chat transcript. It is NOT fetched from a separate endpoint.