Maven/REARCH.md

Maven — Re-architecture (router-centric, 2026-07-10)

Supersedes the classifier-first routing model. Agreed in a design session after diagnosing that homesrv deploys with a stub phraser (no LLM running) and an embedder-classifier that routes by nearest-neighbor between frozen seed phrases — the structural cause of "she messes up queries."

Hardware reality: homesrv = Ryzen 5 5600U laptop, Vega iGPU, 14 GB shared RAM. Workstation (RX 7900 XT) is NOT the deploy target and is often busy. So: small models, on-demand where heavy, always-on where cheap.

Principle

The LLM is not the center of everything. Deterministic tools handle the bulk. The LLM is used for exactly two things: routing/reasoning and talking back. A sub-1B agentic model (LFM 2.5) is enough for both.

If the router is good, Maven feels good. Routing is the linchpin.

The spine

utterance
   → [world-state context]   cheap: time, presence, calendar_busy, weather (no LLM)
   → ROUTER  = LFM (always-on, agentic)
        reads utterance + context + tool schema, emits a STRUCTURED action:
          • call a tool (deterministic)      • answer directly
          • escalate → 4B reasoner (on-demand)
   → tools (deterministic, fast)  /  4B reasoner (on-demand summon)
   → PHRASER = LFM (always-on, same process as router) → TTS / text

• Router = Phraser = one resident sub-1B LFM llama-server, two call-sites (route-prompt, phrase-prompt). Always warm, no cold start. Cheap on 14 GB.
• 4B reasoner (Qwen3-4B, already on disk) — summoned on-demand for genuinely complex turns, torn down / idle-unloaded after. Never resident.
• Embedder demoted from router to tool — it now backs memory.search (RAG) and gives the router a cheap "similar past notes/intents" hint. The router no longer depends on it clearing a threshold. Upgrade MiniLM → bge-m3 for better RU retrieval later (model swap, not architecture).

Router output

• Constrained structured JSON action {tool, args, escalate} — NOT free-form multi-step function-calling. Sub-1B is far more reliable emitting a fixed schema. Enforce with a GBNF grammar in llama.cpp (near-bulletproof).
• Keep the existing stage-0 exact-match fast-path for dead-obvious commands (skips the router entirely) — cheap insurance, already built.

The proactive / memory half — one background engine

"Take notes," "remember," "reflect," "suggest do you want to add X?", "remind" are NOT request-path features. They are one digestion worker:

DIGESTION WORKER  (periodic + event-driven, off the request path)
  • reads new facts/notes since last pass
  • RAG-consolidates: dedupe, link, summarize into durable memory
  • reflects: detect patterns ("mentioned X three times")
  • proposes: "want me to add X / remind you about Y?" → nudge dispatcher
  • surfaces due reminders
  runs LFM (cheap) or summons 4B (real synthesis) — never blocks a turn

Notes capture is a deterministic Tier-0 tool; making notes mean something later is the worker + RAG.

Layer table

Layer	What	Runs
Context	world-state (time/presence/calendar/weather)	always, no LLM
Router	LFM agentic orchestrator — linchpin	always-on
Tools	note/reminder/memory/calendar/weather/act (deterministic)	always
Reasoner	Qwen3-4B for complex turns	on-demand summon
Phraser	LFM — final voice	always-on (same proc as router)
Digestion worker	reflection → suggestions/nudges/memory	background
Reach	telegram (+ existing ntfy/voice)	quick win
Voice quality	custom/better TTS	deferred (workstation GPU busy)

Build order

1. Foundation + router — router-as-LFM, turn the engine ON (resident sub-1B), verify notes+reminders actually round-trip, date/number TTS normalizer, wire telegram reach. After this she's a trustworthy plain assistant.
2. On-demand 4B reasoner — summon/idle lifecycle + router escalation path.
3. Digestion worker — reflection, proactive suggestions, memory consolidation, RAG read-back.
4. Embodiment — voice quality (deferred).

Non-goals (unchanged)

Never phones home. Not a nag. Not autonomous. Feminine-gendered RU self-ref.