Skip to content

Daily Log: 2026-02-04

🧠 Context Summary

Status: 🟢 Stable | Focus: AI Service Stabilization

Today's session was dedicated to stabilizing the Paperless-AI sidecar on starfleet-compute. The service was crashing due to timeouts caused by running a heavy model (mistral-nemo 12B) on the Mac Mini's CPU, along with several configuration gaps in the Docker setup.

🚧 Current Projects

  1. AI Sidecar Stabilization (paperless-ai):

    • Status: ✅ Fixed & Active.
    • Issue 1 (Timeouts): mistral-nemo (12B) was too slow for the application's strict startup validation, causing "Client closed connection" errors.
    • Fix: Switched to qwen2.5:7b. This is the "sweet spot" for CPU inference—significantly faster than Mistral-Nemo but smarter than Llama 3.2 (3B).
    • Issue 2 (Configuration): The service failed with "Failed to get own user ID" and "SyntaxError" for custom fields.
    • Fix: Added PAPERLESS_USERNAME=vivianl and CUSTOM_FIELDS={"custom_fields":[]} to the environment variables.
    • Issue 3 (Provider): The openai provider setting forced a check for gpt-4o-mini, which failed locally.
    • Fix: Switched AI_PROVIDER=custom to bypass hardcoded model checks.
    • Performance: Processing ~16 documents in 2.5 hours. Success verified on Doc IDs 1-13, 15, 17, and 20.

  2. AI Console (Open-WebUI):

    • Status: 🟢 Linked.
    • Host: holodeck-lab (192.168.1.11).
    • Integration: Verified it can see and use the qwen2.5:7b model running on starfleet-compute.

📋 Fleet Health Report (Current)

  • starfleet-compute: 🟢 High Load (Normal). CPU usage ~370% while processing documents. Running ollama + paperless-ai.
  • paperless-ai: 🟢 Processing. Logs confirm "Using character-based truncation for model: qwen2.5:7b".
  • holodeck-lab: 🟢 Online. Open-WebUI is active.

📓 Technical Notes (Critical for Future)

The paperless-ai container MUST have the following environment variables to function in this CPU-based setup:

environment:
  - PAPERLESS_API_URL=http://paperless:8000/api
  - PAPERLESS_API_TOKEN=343c...
  - PAPERLESS_USERNAME=vivianl  # REQUIRED for mapping
  - AI_PROVIDER=custom          # REQUIRED to bypass OpenAI checks
  - CUSTOM_API_KEY=sk-dummy
  - CUSTOM_BASE_URL=http://ollama:11434/v1
  - CUSTOM_MODEL=qwen2.5:7b     # 7B is the max for stability
  - API_KEY=federation_override # For manual API triggers
  - CUSTOM_FIELDS={"custom_fields":[]} # REQUIRED to prevent JSON parse error

💡 Strategic Notes

  • Model Standard: qwen2.5:7b is now the standard fleet model for general instruction following. mistral-nemo and llama3.2 have been purged to save space.
  • Performance: Processing a document takes ~5-10 minutes on the i7 CPU. This is expected. Do not assume "hang" unless logs are silent for >15 minutes.