From 3babac58699110dcf385cf23c5b24470f404e9f4 Mon Sep 17 00:00:00 2001
From: jaewilson07 <jaewilson07@gmail.com>
Date: Fri, 24 Apr 2026 08:31:44 -0600
Subject: [PATCH] refactor(voice-to-mermaid): remove vtm-ollama sidecar,
 default to qwen3:8b
MIME-Version: 1.0
Content-Type: text/plain; charset=UTF-8
Content-Transfer-Encoding: 8bit

Consolidates on Windows-native Ollama via host.docker.internal:11434:
- Remove vtm-ollama sidecar container + ollama-data volume (was vestigial —
  api was already pointing at host.docker.internal by default)
- Update default_model from gemma4:31b (did not exist) to qwen3:8b
- Update model_filter to match models actually installed on Windows Ollama

Part of the 2026-04-24 reproducible-infra consolidation — full context at
simpleDiscordBot/.agents/plans/2026-04-24_reproducible-infra-consolidation/

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
---
 voice-to-mermaid/backend/config.yaml        | 17 +++++---
 voice-to-mermaid/backend/docker-compose.yml | 43 ++++++++++-----------
 2 files changed, 33 insertions(+), 27 deletions(-)

diff --git a/voice-to-mermaid/backend/config.yaml b/voice-to-mermaid/backend/config.yaml
index 936fb6d..e58a2fb 100644
--- a/voice-to-mermaid/backend/config.yaml
+++ b/voice-to-mermaid/backend/config.yaml
@@ -4,15 +4,22 @@
 # Secrets (OLLAMA_URL, OPENAI_API_KEY, etc.) stay in .env / Infisical.
 
 ollama:
+  # Default model for first-diagram generation. qwen3:8b is chosen for:
+  # - Fast first-token latency (important for "type and see diagram" UX)
+  # - Good instruction-following on mermaid grammar
+  # - Fits in 8-12 GB VRAM (cold-load ~5s, then streams)
+  # Swap to a 32B class model (qwen2.5-coder:32b, qwen3.5:27b) for complex
+  # multi-node diagrams at the cost of ~20s cold-load latency.
   default_model: "qwen3:8b"
-  # Models shown in the UI picker — exact names or prefixes (e.g. "qwen2.5:").
-  # Leave empty to show all installed models.
+  # Models shown in the UI picker. Must exist in Windows Ollama
+  # (C:\Users\jaewi\.ollama is not used — see OLLAMA_MODELS=D:\ollama\models).
+  # Update as needed with: ollama pull <model>  (run in Windows PowerShell)
   model_filter:
+    - "qwen3:8b"
     - "qwen2.5:14b"
+    - "qwen2.5-coder:32b"
     - "qwen3.5:27b"
-    - "qwen3:8b"
-    - "qwen3:14b"
-    - "llama3.2:3b"
+    - "gemma3:12b"
 
 openai:
   model: "gpt-4o-mini"
diff --git a/voice-to-mermaid/backend/docker-compose.yml b/voice-to-mermaid/backend/docker-compose.yml
index d1601d7..7fe68da 100644
--- a/voice-to-mermaid/backend/docker-compose.yml
+++ b/voice-to-mermaid/backend/docker-compose.yml
@@ -1,21 +1,18 @@
-services:
-  ollama:
-    image: ollama/ollama:latest
-    container_name: vtm-ollama
-    restart: unless-stopped
-    ports:
-      - "11434:11434"
-    volumes:
-      - ollama-data:/root/.ollama
-    # GPU support — remove 'deploy' block if no GPU available
-    deploy:
-      resources:
-        reservations:
-          devices:
-            - driver: nvidia
-              count: 1
-              capabilities: [gpu]
+# Voice-to-Mermaid — LOCAL MACHINE ONLY (requires GPU)
+# ⚠️  GPU GUARD: This compose file requires NVIDIA GPU hardware.
+# ⚠️  DO NOT deploy to the VPS (187.77.216.108) — it has no GPU.
+# ⚠️  See infrastructure/local/README.md for local machine service registry.
 
+#
+# Architecture (as of 2026-04-24):
+#   vtm-api → http://host.docker.internal:11434 → Windows-native Ollama
+#
+# The previous `vtm-ollama` sidecar + `ollama-data` volume were removed —
+# they duplicated the Windows Ollama instance (ran empty, wasted resources).
+# See .agents/runbooks/troubleshoot-ollama/SKILL.md for the single-instance
+# architecture rules.
+#
+services:
   api:
     build: .
     container_name: vtm-api
@@ -24,14 +21,16 @@ services:
       - "${PORT:-7625}:7625"
     env_file: .env
     environment:
-      # Default to the bundled ollama service; override in .env to use a
-      # host-resident Ollama (http://host.docker.internal:11434) instead.
-      OLLAMA_URL: ${OLLAMA_URL:-http://ollama:11434}
+      # Windows-native Ollama via Docker Desktop gateway.
+      OLLAMA_URL: ${OLLAMA_URL:-http://host.docker.internal:11434}
     extra_hosts:
       - "host.docker.internal:host-gateway"
     volumes:
       - ./prompts:/app/prompts       # mount prompts so edits don't need rebuild
       - ./config.yaml:/app/config.yaml  # mount config so edits don't need rebuild
+    networks:
+      - local-ai
 
-volumes:
-  ollama-data:
+networks:
+  local-ai:
+    name: local-ai