[Release] Nexidion – A private knowledge vault with an autonomous local AI background worker.

Release

After almost two years of development and several architectural revisions, I’m finally releasing a project that addresses my privacy concerns: Nexidion. This is an open-source hierarchical Markdown note-taking app with a built-in, optional autonomous background worker designed to work seamlessly with local OpenAI-compatible models such as llama.cpp, Ollama, and others.

The Local LLM Angle: Autonomous Background Worker

Nexidion is a note-taking app where instead of just chatting with your notes, you can instruct the background worker to perform tasks like organizing messy notes into hierarchical folders, summarizing subtrees, or extracting action items from meeting notes. The worker operates in the background and commits changes under its own name within a version-controlled system. This ensures that if anything goes wrong, you have full control over reverting any alterations.

The safety net: To prevent errors like hallucinations by the AI, Nexidion includes a robust version control mechanism. If your local model botches an organization task, you can revert it with just one click without losing any data or making external calls to cloud services. This setup is ideal for those with more modest hardware like my single RTX 2080 Ti GPU.

My "GPU Poor" Setup (2080 Ti)

You don’t need a high-end multi-GPU setup to leverage the background worker. I’m running this on a single RTX 2080 Ti with 11GB VRAM, and it works surprisingly well for tasks like organizing notes. Here’s how you can set up a similar environment:

bash docker run --gpus all --rm \
-v /mnt/c/.../models/unsloth/Qwen3.5-36B:/models \
havenoammo/llama:cuda12-server \
-m /models/Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf \
--port 1234 --host 0.0.0.0 \
-n -1 --parallel 1 --threads 6 \
--ctx-size 100000 --fit-target 844 \
--mmap -ngl 18 --flash-attn on \
--temp 1.0 --min-p 0.0 --top-p 0.95 --top-k 20 \
--jinja \
--chat-template-kwargs '{"preserve_thinking":true}' \
--ubatch-size 512 --batch-size 2048 \
--cache-type-k q8_0 --cache-type-v q8_0 \
--spec-type ngram-mod,draft-mtp \
--spec-draft-n-max 3

I’ve Dockerized everything, so setting up the Postgres DB, backend, frontend, and the AI task runner can be done with a single command:

bash docker compose --profile with-postgres --profile with-task-runner up -d

(Full documentation and setup instructions are available in the repo).

I’d love to hear your feedback! Specifically, I’m interested in hearing from those familiar with local AI models on how the background agent performs with different quantizations and what specific prompts work best for organizing notes. Please let me know what you think!

Key Takeaways

• Nexidion is a private knowledge vault with an autonomous background worker designed to operate locally.
• The app uses version control to ensure data safety, even when the AI makes mistakes.
• You can run Nexidion on modest hardware like a single RTX 2080 Ti GPU.

Stay ahead of AI. Get the most important stories delivered to your inbox — no spam, no noise.

Please enable JavaScript in your browser to complete this form.

Name *

First

Last

Email Name

Email *

Submit