Searching my own documents with RAG
A search service over my own notes and documents, built to run in the memory a shared server has.
- Status
- Live
- Role
- Everything
- Period
- 2026, ongoing
- Built with
- Python, fastembed, sqlite-vec, MCP
- Scope
- Just over 1,600 documents
The starting point
An archive of notes is only useful if you can find what you wrote six months ago, even when you do not remember the words you used.
The choices
Everything had to run alongside other services on one server, without a GPU, a large vector database or a per-query fee. So:
- Small embeddings from fastembed
- Vectors stored in SQLite with sqlite-vec
- Search exposed over MCP, so AI tools can look things up in the notes
Where it stands
The service runs permanently on my own server and indexes just over 1,600 documents in more than 20,000 chunks. The quality of the answers still depends on how well a document is written.
The main lesson
How a document is split into chunks mattered more to search quality than the choice of model.
Should your own documents be searchable?
That is a job SES Group can take on.