All work
Client work · Private AI
100% offline RAG over institutional memory
Client: a mid-size engineering firm · Constraint: no cloud AI
01
Problem
Leadership couldn't query years of institutional memory scattered across mailboxes and documents — and cloud AI was off the table for confidentiality reasons.
02
Approach
A 100% local RAG stack: Ollama embeddings, SQLite with sqlite-vec and FTS5, and RRF hybrid search combining vector and full-text retrieval.
An idempotent daily sync keeps the index current without re-processing anything that hasn't changed.
The search core is open source.
github.com/giuseppeferretti/sqlite-rag-mcp
03
Results
- emails ingested, plus 34 documents
- 3,047
- emails ingested, plus 34 documents
- chunks indexed
- 15,264
- chunks indexed
- structured memories extracted
- 4,498
- structured memories extracted
- cloud exposure — everything runs on-premises
- 0
- cloud exposure — everything runs on-premises
Published with the client's written authorization; identifying details withheld.
Want results like these?
A 20-minute call is enough to tell whether your process can be automated — and what it would take.