# Architecture — Plane Mode Scholar

## Agent pipeline (Best Agent)

```
User message
  → classify_intent
  → retrieve_chunks (vector search on pack)
  → retrieve_memories (hybrid filter + rerank)
  → pack_prompt (budget: chunks + memories + working memory)
  → generate (Qwen3.5-9B, streaming)
  → extract_candidates (regex + LLM JSON)
  → persist / consolidate / index_memory
```

## Memory types

| Type | Written when |
|------|----------------|
| misconception | Quiz miss, teach-back gap, user states confusion |
| review_item | SRS scheduler after miss or gap |
| session_summary | End session |
| study_goal / deadline | Pack create or chat extraction |
| user_preference | Chat patterns |

## Spaced repetition

SM-2 lite in `memory/scheduler.py`. First review is **due immediately** after a quiz miss so the Home dashboard and recall banner work within the same demo session.

## Storage

- SQLite: packs, sessions, turns, memories, logs
- Vector stores: `pack_{id}` for chunks, `memory_{pack_id}` for memories
- Persistent disk on HF Space: `/data/plane_mode_scholar`

## Judge automation

- `docs/readiness_cases.jsonl` — 5 scripted scenarios
- `scripts/readiness_check.py` — automated SRS + trace shape checks
- `tests/test_scheduler.py` — unit tests for due-now scheduling