# NODA2 Capabilities — Machine-readable for NODA1 Router
# Generated: 2026-02-27
# Node: MacBook Pro M4 Max, 64GB RAM, macOS 26.3

node_id: noda2
hostname: MacBook-Pro.local
nats_endpoint: "nats://145.224.111.147:4222"  # leafnode spoke, via NODA1 hub
router_url: "http://145.224.111.147:9102"      # NODA2 router (external, if firewall allows)
swapper_url: "http://145.224.111.147:8890"     # swapper (external, if firewall allows)
ollama_url: "http://145.224.111.147:11434"     # Ollama GPU (external, if firewall allows)

# ── Vision Models ──────────────────────────────────────────────────────────
vision_models:
  - name: llava:13b
    backend: ollama
    base_url: http://localhost:11434
    type: vision
    size_gb: 8.0
    quality: medium
    note: "LLaVA 1.6 + CLIP, 13B. Available now. Legacy quality."
    capability_hint: "image_understanding"

  # RECOMMENDED - not yet installed, install with: ollama pull qwen3-vl:8b
  - name: qwen3-vl:8b
    backend: ollama
    base_url: http://localhost:11434
    type: vision
    size_gb: ~8.0
    quality: high
    installed: false
    note: "Install: ollama pull qwen3-vl:8b. Superior OCR, document, chart understanding."

# ── Text/LLM Models ────────────────────────────────────────────────────────
text_models:
  - name: qwen3.5:35b-a3b
    backend: ollama
    base_url: http://localhost:11434
    type: llm
    size_gb: 9.3
    params: "14.8B active (MoE 35B total)"
    quality: high
    speed: fast
    note: "Primary recommendation: MoE architecture, fast inference on M4 Max"

  - name: qwen3:14b
    backend: ollama
    base_url: http://localhost:11434
    type: llm
    size_gb: 9.3
    params: "14B"
    quality: high
    speed: medium

  - name: glm-4.7-flash:32k
    backend: ollama
    base_url: http://localhost:11434
    type: llm
    size_gb: 19.0
    params: "32B"
    quality: high
    speed: slow
    note: "Long context (32k). Use for document-heavy tasks."

  - name: deepseek-r1:70b
    backend: ollama
    base_url: http://localhost:11434
    type: llm
    size_gb: 42.0
    params: "70B"
    quality: very_high
    speed: very_slow
    note: "Reasoning model. Use for complex multi-step planning only."

  - name: Qwen3.5-35B-A3B-Q4_K_M
    backend: llama-server
    base_url: http://localhost:11435
    api_type: openai_compatible
    type: llm
    note: "Same model as qwen3.5:35b-a3b via llama.cpp. Redundant with Ollama."

# ── Code Models ────────────────────────────────────────────────────────────
code_models:
  - name: deepseek-coder:33b
    backend: ollama
    base_url: http://localhost:11434
    type: code
    size_gb: 18.0

  - name: starcoder2:3b
    backend: ollama
    base_url: http://localhost:11434
    type: code
    size_gb: 1.7
    speed: very_fast

# ── Embedding Models ───────────────────────────────────────────────────────
embedding_models:
  # Not explicitly found in Ollama list
  # memory-service uses its own embedding (check dagi-memory-service-node2)
  - name: unknown
    note: "Audit needed: check memory-service embedding model config"

# ── STT Models ─────────────────────────────────────────────────────────────
stt_models:
  - name: whisper (swapper)
    backend: swapper
    base_url: http://localhost:8890
    endpoint: /stt
    note: "Swapper exposes /stt/models (HTTP 200). Model details need swapper config check."

# ── TTS Models ─────────────────────────────────────────────────────────────
tts_models:
  - name: tts (swapper)
    backend: swapper
    base_url: http://localhost:8890
    endpoint: /tts
    note: "Swapper exposes /tts/models (HTTP 200). Model details need swapper config check."

# ── Routing Policy ─────────────────────────────────────────────────────────
routing_policy:
  vision_request:
    capability: "vision"
    selection_logic: |
      if quality_tier == "best":
        prefer: qwen3-vl:8b (when installed)
        fallback: llava:13b
      if quality_tier == "fast":
        prefer: llava:13b
    return_model_used: true
    nats_subject: "node.noda2.vision.request"

  text_request:
    capability: "text"
    selection_logic: |
      if quality_tier == "best": deepseek-r1:70b
      if quality_tier == "high": qwen3.5:35b-a3b (default)
      if quality_tier == "fast": qwen3:14b or gemma3:latest
    nats_subject: "node.noda2.llm.request"

  code_request:
    capability: "code"
    selection_logic: |
      prefer: deepseek-coder:33b
      fast fallback: starcoder2:3b
    nats_subject: "node.noda2.code.request"