We’ve been building NabuAI – an AI-powered knowledge management tool for researchers and knowledge workers. Here are a few lessons from the first months.

1. RAG before custom models

We tried to get fancy with custom embeddings and fine-tuning early. It slowed us down and didn’t move the needle for our early users. Switching to a simple GraphRAG-style pipeline (chunk → embed → retrieve → generate) gave us:

• Faster iteration
• Good enough quality for v1
• Room to improve retrieval and prompts without changing the whole stack

So: get RAG right first, then consider custom models only when you have clear, measurable gaps.

2. Where time actually went

Rough breakdown of where engineering time went in the first 3 months:

Integrations and UI took more than we’d guessed; infra stayed manageable thanks to Supabase and Vercel.

3. Links we leaned on

• LlamaIndex docs – for RAG patterns and graph-based retrieval
• Vue 3 + TypeScript – for the frontend
• Tailwind CSS – for layout and theming

4. One code decision that helped

We kept “sources” for every answer in a simple table and exposed them in the UI from day one:

answer_sources: [ document_id, chunk_id, score ]

That made it easy to show “from which note this came” and to debug bad answers. Small schema choice, big product and debugging win.

More posts on GraphRAG, the Chrome extension, and our stack choices are planned. If you’re building something similar, say hi.