Mechanics Sidekick

Everything runs against Ollama on my own machine. No API keys, no cloud, no monthly bill. The honest reason is I wanted to play with local LLMs and see what the current generation could actually do. Nothing leaving my machine is a nice side effect, not the driver.

Most RAG setups reach for pgvector, Qdrant, or Chroma. I just stored everything in SQLite, embeddings and all. This was my first RAG project, it's single-user, and I wanted to build it with tools I already knew rather than fighting a new database at the same time as learning RAG. Swapping in a real vector store later is straightforward if I ever need it, but for a single-user CLI SQLite is fine.

Naive chunking on service manuals destroys context — a torque value lifted from the middle of a procedure loses the engine variant, the system, and the step it belonged to. So before each chunk gets embedded, a lightweight LLM call generates a short summary that situates it (which engine, which system, which procedure) and that summary is prepended to the chunk text. It's contextual retrieval — the embedding represents the chunk plus its place in the manual, not just the raw words.

Chunking itself is structure-aware rather than fixed-size. The PDF extractor reads font metadata (bold, ALL CAPS, font-size jumps) to detect section headings, and every chunk carries the section title it came from. That keeps related steps together and gives the retriever something to anchor on beyond raw similarity.