🛠️ MyZubster Dev Update — Local AI + RAG
Today we made an important step forward in MyZubster’s local AI infrastructure.
We now have a local stack built around:
MyZubster API → Qdrant → Ollama → Mistral
with nomic-embed-text generating embeddings and Mistral running locally to generate answers.
🔗 MyZubster MVP
https://github.com/nicolaususnicola-lgtm/myzubster-mvp
🔗 Ollama
https://ollama.com/
🔗 Qdrant
https://qdrant.tech/
What changed?
The main work today was on the RAG — Retrieval-Augmented Generation — pipeline.
Previously, every AI question caused MyZubster to regenerate embeddings for all observations and upsert them into Qdrant again before searching.
We separated indexing from retrieval.
Now:
When an observation is recorded
→ MyZubster stores it
→ its embedding is generated
→ it is indexed in Qdrant
When a question is asked
→ only the question embedding is generated
→ Qdrant retrieves the most relevant information
→ that context is passed to Mistral
→ Mistral generates the answer
This avoids repeatedly indexing the same observations and moves MyZubster toward a persistent, searchable local knowledge layer.
We also worked on the Kefir & Knowledge / KF-006 data and made an important evidence distinction explicit:
RECORDED = digitally recorded inside MyZubster
onchainRecorded=false = no blockchain registration is being claimed
That distinction reflects one of the principles we are applying throughout development:
Evidence first. Claim only what the available data can actually verify.
The backend refactor was validated with the automated test suite:
✅ 25/25 tests passing
Next step
The next stage is to connect the broader MyZubster knowledge base — including N4K48, Zorgax, roadmap and project documentation — to the same RAG architecture, with a consistent data schema, deterministic indexing and duplicate control.
The goal is not simply to build another chatbot.
The goal is to build a local MyZubster AI that can retrieve and reason over the project’s own knowledge and evidence.
Top comments (0)