Rebalance homelab AI routing around NOMAD-first inference

DEPLOYMENT_HANDOFF.md

@@ -7,6 +7,7 @@ Phase 2 (AI stack) is fully deployed and verified on PlausibleDeniability (10.5.
 ## What's Done
 
 ### Phase 2 — AI Stack (COMPLETE)
+Historical note: this handoff reflects the original PD-first AI deployment snapshot, not the current Honcho routing policy.
 All 6 containers running and verified on PD (10.5.30.6):
 - **Ollama:** port 11434 — qwen2.5:14b loaded
 - **OpenWebUI:** port 8282 (→8080)

HOMELAB_BUILDOUT_PLAN.md

@@ -575,7 +575,7 @@
    575|Trigger:  Cron node (every Sunday at 10:00 AM)
    576|Step 1:   Postgres node → SELECT watch history for past 7 days
    577|Step 2:   Format data as structured text
-   578|Step 3:   HTTP Request → Ollama API (POST to PD's qwen2.5:14b)
+   578|Step 3:   HTTP Request → Ollama API (POST to the local/default automation model host, preferably N.O.M.A.D.-local rather than PD)
    579|            - Prompt: "Summarize this week's viewing in a fun digest: {data}"
    580|Step 4:   HTTP Request → Gotify (send digest)
    581|```
@@ -1093,9 +1093,9 @@
   1388|
   1389|LiteLLM sits in front of your three Ollama instances and presents a single OpenAI-compatible API at `http://litellm:4000`. It handles:
   1390|
-  1391|- **Model routing** — requests for `qwen3:32b` go to ROCINANTE, `qwen2.5:14b` goes to PD, `phi4` goes to N.O.M.A.D.
+  - **Model routing** — current reliable default is N.O.M.A.D.-local Ollama for background/automation paths, with PD kept available for shared light-tier use and Rocinante treated as optional/manual heavy capacity.
-  1392|- **Failover** — if ROCINANTE is offline, LiteLLM can fall back to PD automatically
+  - **Failover** — avoid designs that require Rocinante to be online; degrade to N.O.M.A.D.-local or explicitly operator-invoked PD paths instead of automatic dependence on a personal PC.
-  1393|- **Load balancing** — distribute requests across instances running the same model
+  - **Load balancing** — distribute requests across instances running the same model when you have multiple dependable backends, but do not assume Rocinante qualifies as one.
   1394|- **Usage tracking** — logs token counts, latency, and costs per model/user via its built-in database
   1395|
   1396|### 7.2 — Scaffold Directories
@@ -1621,17 +1621,23 @@
   1916|{
   1917|  "models": [
   1918|    {
-  1919|      "title": "PD - qwen2.5:14b",
+  1919|      "title": "N.O.M.A.D. - qwen2.5:3b",
   1920|      "provider": "ollama",
-  1921|      "model": "qwen2.5:14b",
+  1921|      "model": "qwen2.5:3b",
-  1922|      "apiBase": "http://PD_TAILSCALE_IP:11434"
+  1922|      "apiBase": "http://NOMAD_LOCAL_OR_TAILSCALE_IP:11434"
   1923|    },
   1924|    {
-  1925|      "title": "ROCINANTE - qwen3:32b",
+  1925|      "title": "PD - shared light tier (manual / fallback)",
   1926|      "provider": "ollama",
-  1927|      "model": "qwen3:32b",
+  1927|      "model": "qwen2.5:14b",
-  1928|      "apiBase": "http://ROCINANTE_TAILSCALE_IP:11434"
+  1928|      "apiBase": "http://PD_TAILSCALE_IP:11434"
-  1929|    }
+  1929|    },
+  1930|    {
+  1931|      "title": "ROCINANTE - opportunistic heavy",
+  1932|      "provider": "ollama",
+  1933|      "model": "qwen3:32b",
+  1934|      "apiBase": "http://ROCINANTE_TAILSCALE_IP:11434"
+  1935|    }
   1930|  ],
   1931|  "tabAutocompleteModel": {
   1932|    "title": "N.O.M.A.D. - phi4",

docs/architecture/ARCHITECTURE_OVERVIEW.md

@@ -40,7 +40,8 @@ Current live homelab stack plus the planned direction for the upgraded PD platfo
 - **Dockhand** (3230) — Docker management UI
 
 ### AI Layer
-- **LiteLLM** (4000) — unified LLM proxy; routes to Ollama on PD (light) and Rocinante (heavy)
+- **LiteLLM** (4000) — unified LLM proxy; PD remains the shared light tier, while Rocinante is an opportunistic heavy tier when the PC is online
+- **Honcho** — memory/deriver stack on N.O.M.A.D. using local Ollama first so PD outages do not cascade into MeshMonitor/database incidents
 - **OpenWebUI** (8282) — chat UI with RAG, STT, and web search
 - **Qdrant** (6333/6334) — vector database for RAG
 - **Whisper** (8786 / 8787) — CPU speech-to-text on N.O.M.A.D. plus CUDA speech-to-text on Rocinante

docs/architecture/SERVICES_DIRECTORY.md

@@ -40,8 +40,8 @@ All homelab services, their hosts, ports, and current status.
 | Whisper (faster-whisper, CPU) | N.O.M.A.D. (10.5.30.7) | 8786 | ✅ Active |
 | Whisper (faster-whisper, CUDA large-v3) | Rocinante (10.5.30.112) | 8787 | ⚠️ Re-check from N.O.M.A.D. if Whisper-specific traffic is needed; Ollama-only recovery was validated 2026-05-27 |
 | SearXNG | PD | 8888 | ✅ Active |
-| Ollama — light tier | PD (10.5.30.6) | 11434 | ✅ Active |
+| Ollama — light tier | PD (10.5.30.6) | 11434 | ✅ Active; keep routine Honcho load off this host to protect MeshMonitor/shared Postgres |
-| Ollama — heavy tier | Rocinante (10.5.30.112) | 11434 | ✅ Reachable from N.O.M.A.D.; re-validated 2026-05-27 |
+| Ollama — heavy tier | Rocinante (10.5.30.112) | 11434 | ⚠️ Personal PC / opportunistic only; do not depend on it for routine background routing |
 | Reranker (TEI) | Serenity (10.5.30.5) | 9787 | ✅ Active |
 
 OpenClaw is no longer treated as an active PD service in the operator inventory. Only helper-artifact references remain under the school-intake planning docs.

docs/servers/PLAUSIBLEDENABILITY.md

@@ -47,9 +47,10 @@ Recommended models for 11GB VRAM:
 | `phi4` | ~8GB | Analytical tasks |
 
 Current light-tier notes:
-- PD Ollama is the active light-tier endpoint for LiteLLM on `http://10.5.30.6:11434`
+- PD Ollama remains the shared light-tier endpoint for the broader AI stack on `http://10.5.30.6:11434`, but it should be protected from routine Honcho background load.
-- Local Honcho currently points at PD Ollama over LAN (`http://10.5.30.6:11434/v1`) instead of using a separate local model host
+- Local Honcho now uses N.O.M.A.D.'s own Ollama on `http://127.0.0.1:11434/v1` so PD pressure does not take down MeshMonitor or shared Postgres.
-- Current Honcho routing split keeps embeddings on `nomic-embed-text:latest`, cheap/default paths on `qwen2.5:7b`, and deeper high/max paths on `qwen2.5:14b`
+- Current Honcho routing keeps embeddings on `nomic-embed-text:v1.5`, cheap/default paths on `qwen2.5:3b`, and high/max paths on `llama3.1:latest`.
+- Rocinante is John's PC and should be treated as an opportunistic/manual heavy tier, not a dependable always-on backend.
 
 ## Remote Access
 - Pangolin VPS + Newt for external services

docs/servers/ROCINANTE.md

@@ -24,15 +24,15 @@ Ollama runs on port `11434` and hosts the large-model ("heavy") tier.
 | `qwen2.5:32b` | Large-context tasks |
 | `gemma3:27b` | Multimodal / vision |
 
-> **LiteLLM tier:** Rocinante is the **"heavy"** tier in LiteLLM `config.yaml`.
+> **LiteLLM tier:** Rocinante may serve as an **opportunistic heavy/manual tier** when the PC is online.
-> Heavy-tier requests from OpenWebUI and n8n route to `http://10.5.30.112:11434`.
+> Do not depend on `http://10.5.30.112:11434` for routine background routing; verify live reachability first and prefer N.O.M.A.D.-local paths for guaranteed availability.
 
 ## Current operational note
 
-- VLAN access to Rocinante from N.O.M.A.D. was re-validated on 2026-05-27.
+- Rocinante is John's personal PC and is not a dependable always-on homelab tier.
-- Rocinante Ollama now answers from N.O.M.A.D. at `http://10.5.30.112:11434`; `/api/tags`, `/api/generate`, and OpenAI-compatible `/v1/chat/completions` all succeeded.
+- Treat Ollama/Whisper on Rocinante as opportunistic capacity when the machine is online, not as a required background dependency.
-- Hermes configs that still pointed at the legacy `10.5.1.112` endpoint were corrected to `10.5.30.112`.
+- If N.O.M.A.D. or PD need guaranteed inference, prefer local N.O.M.A.D. routing first and keep PD protected from routine background load.
-- Re-check CUDA Whisper on `http://10.5.30.112:8787/health` separately if Whisper-specific reachability is needed; this pass validated Ollama end-to-end.
+- Re-check live reachability from N.O.M.A.D. before any Rocinante-dependent maintenance or cutover.
 
 ## Notes
 

n8n-workflows/README.md

@@ -171,10 +171,16 @@ curl -X PUT http://10.5.30.6:6333/collections/knowledge_base \
 
 (768 dimensions matches `nomic-embed-text`. If you use a different embedding model, adjust the size.)
 
-### Pull nomic-embed-text into Ollama (required for workflows 02, 05, 08)
+### Pull `nomic-embed-text` into the Ollama host backing these workflows (required for workflows 02, 05, 08)
 
+For N.O.M.A.D.-local workflows / Honcho-style routing:
 ```bash
-curl http://10.5.30.6:11434/api/pull -d '{"name": "nomic-embed-text"}'
+curl http://127.0.0.1:11434/api/pull -d '{"name": "nomic-embed-text:v1.5"}'
+```
+
+For PD-hosted shared workflows, use the PD endpoint explicitly instead:
+```bash
+curl http://10.5.30.6:11434/api/pull -d '{"name": "nomic-embed-text:latest"}'
 ```
 
 ### Paperless Webhook (required for workflows 03, 05, 17, and 19)