BITE uses a local-first three-layer architecture.
┌─────────────────────────────────────────────────────────┐
│ Output layer obsidian-vault/ideas/ │
│ ideas, plans, review notes │
├─────────────────────────────────────────────────────────┤
│ Index layer obsidian-vault/index/ │
│ JSONL index + navigation pages │
├─────────────────────────────────────────────────────────┤
│ Evidence layer obsidian-vault/analysis/ │
│ structured notes + logs │
│ obsidian-vault/paperPDFs/ │
│ source PDFs │
└─────────────────────────────────────────────────────────┘
obsidian-vault/paperPDFs/stores source PDFs.obsidian-vault/analysis/stores per-paper structured analysis notes and logs.- This is the primary evidence surface that agents should read.
obsidian-vault/index/stores generated retrieval indexes and navigation pages.- At scale, agents should start from
index.jsonl, filter candidates, then read matching analysis notes.
obsidian-vault/ideas/stores brainstorm notes, focused plans, reviewer critiques, and daily logs.- This layer is downstream of evidence and index, not a substitute for them.
When the vault is large, use:
index.jsonl -> filter -> matching analysis notes -> downstream ideas