diff --git a/Asset-Label-Matching.md b/Asset-Label-Matching.md
index ec1fdd0..1b2df68 100644
--- a/Asset-Label-Matching.md
+++ b/Asset-Label-Matching.md
@@ -1,6 +1,6 @@
 # Asset Label Photo Matching
 
-**Added:** 2026-05-29
+**Updated:** 2026-05-29
 
 New system for matching label photos (serial numbers, barcodes, QR codes) against the assets database.
 
@@ -21,10 +21,17 @@ Examples:
 
 `find_asset_by_normalized_id()` searches across `serial_number`, `connect_id`, `equipment_id`, `machine_id`, and `barcode` columns — normalizing all before comparing.
 
+## OCR Pipeline (priority order)
+
+1. **Ollama vision model** (qwen2.5vl:3b) — runs on Windows gaming PC (192.168.0.181) via persistent SSH tunnel (systemd `ollama-tunnel.service` on localhost:11434). Most accurate. Resizes images to 640x480 before sending.
+2. **Tesseract** (pytesseract) — local fallback, only used if Ollama is unavailable.
+
+Response includes `ocr_source` field (`"ollama"` or `"tesseract"`).
+
 ## API Endpoints
 
 ### `POST /api/ocr`
-Existing photo upload endpoint enhanced. Now returns `matched_assets` in addition to the legacy `machine_id` field. Uses Tesseract OCR. Falls through gracefully when OCR produces garbage.
+Existing photo upload endpoint enhanced. Now returns `matched_assets` and `ocr_source` in addition to the legacy `machine_id` field. Ollama takes priority over Tesseract.
 
 ### `POST /api/match-text`
 New endpoint for client-side processing:
@@ -41,6 +48,18 @@ python3 scripts/match_label_photo.py photo.jpg
 python3 scripts/match_label_photo.py --text "S/N: 2500.0100.0025534"
 ```
 
-## Vision Fallback
+## Windows PC Setup
 
-The vision model at opencode.ai (`mimo-v2-omni`) is used when Tesseract fails on dark-background or complex labels. The app can accept text from client-side vision processing via `/api/match-text`.
+- **Host:** gamingpc (100.84.53.121 / 192.168.0.181)
+- **Service:** Ollama with qwen2.5vl:3b model
+- **Connection:** SSH tunnel via systemd `ollama-tunnel.service`
+- **Auth:** SSH key at `~/.ssh/id_comfyui`
+- **Tunnel:** `ssh -N -L 11434:127.0.0.1:11434 gamingpc`
+- **Auto-start:** `sudo systemctl enable ollama-tunnel.service`
+
+Results with qwen2.5vl:3b on test photos:
+| Photo | Raw Text | DB Match |
+|-------|----------|----------|
+| Keurig label | `S/N: 2500.0100.0025534` | ✅ Asset #5144 (636671 / 956 Cypress Way) |
+| Coca-Cola cooler | `2010378A00039 / RCUCC095.6` | ❌ Not in DB (bottler tags) |
+| Telemetry device | `ID# 44343331624226353 / Monyx ID 48602143` | ❌ Not in DB (GPS tracker) |
diff --git a/Ollama-Vision-Tunnel.md b/Ollama-Vision-Tunnel.md
new file mode 100644
index 0000000..93e6728
--- /dev/null
+++ b/Ollama-Vision-Tunnel.md
@@ -0,0 +1,29 @@
+# Ollama Vision Tunnel Service
+
+## Install
+
+```bash
+sudo tee /etc/systemd/system/ollama-tunnel.service << 'EOF'
+[Unit]
+Description=SSH Tunnel to Windows PC Ollama
+After=network-online.target
+Wants=network-online.target
+
+[Service]
+Type=simple
+User=oplabs
+ExecStart=/usr/bin/ssh -o ConnectTimeout=5 -o ServerAliveInterval=30 -o ServerAliveCountMax=3 -o StrictHostKeyChecking=no -o ExitOnForwardFailure=yes -N -L 11434:127.0.0.1:11434 gamingpc
+ExecStop=/usr/bin/kill $MAINPID
+Restart=always
+RestartSec=10
+StartLimitIntervalSec=60
+StartLimitBurst=5
+
+[Install]
+WantedBy=multi-user.target
+EOF
+
+sudo systemctl daemon-reload
+sudo systemctl enable ollama-tunnel.service
+sudo systemctl start ollama-tunnel.service
+```