🏦 Local Bank Statement Assistant (Offline LLM)

Upload a bank statement (CSV / Excel / PDF), store it locally in SQLite, and ask questions about it in plain language. The language model runs fully on your machine via node-llama-cpp — there is no Ollama, no external server, and no cloud. After you download a model once, the app works completely offline and your bank data never leaves your device.

Works on macOS and Windows (and Linux). No C/C++ compiler required — both the LLM engine (prebuilt llama.cpp binaries) and SQLite (Node's built-in node:sqlite) ship ready to run.

Requirements

• Node.js ≥ 22.5 (uses the built-in node:sqlite). Check with node --version. Download from https://nodejs.org.

Install & run

npm install
npm start

Then open http://localhost:3000 in your browser.

How to use

1. Download a model — on first launch, pick a small model and click Download. It downloads once from Hugging Face. After that you'll see the "This app is now offline" notice and never need the internet again.
   • Llama 3.2 1B (default) — good balance, ~0.8 GB.
   • Liquid AI LFM2 1.2B — fast edge model, ~0.8 GB.
   • Liquid AI LFM2 2.6B — better reasoning, ~1.7 GB.
   • Qwen2.5 1.5B — strong small model, ~1 GB.
   • Llama 3.2 3B — best for number reasoning, ~2 GB.
   • Qwen2.5 0.5B — tiniest/fastest, weakest at math.
2. Upload your bank statement — CSV, .xlsx/.xls, or a text-based PDF. It's parsed and stored in a local SQLite database (data/bank.db).
3. Ask anything — e.g. "How much did I spend in total?", "What was my biggest debit?", "List all grocery transactions."

Tip: for accurate totals and arithmetic, use the 1B or 3B model. The 0.5B model is fast but unreliable with numbers.

How it works

Concern	Implementation
LLM inference	node-llama-cpp (embedded llama.cpp), streamed token-by-token
Model download	node-llama-cpp model downloader, from Hugging Face, into models/
Storage	Built-in node:sqlite with FTS5 (data/bank.db)
File parsing	papaparse (CSV), xlsx (Excel), pdf-parse (PDF)
Retrieval	Whole statement fed to the model when it fits; FTS5 keyword search otherwise
UI	Express server + a single static HTML/JS page

Notes

• Models and your data are stored under models/ and data/ and are git-ignored — they never get committed.
• This replaces the previous Python/Streamlit + Ollama prototype (app.py), which required a separately-installed Ollama server.

Project layout

server.js          Express server + API routes
src/
  models.js        Catalog of downloadable models
  llm.js           node-llama-cpp: download, load, streamed chat
  db.js            node:sqlite storage + FTS5 search
  ingest.js        CSV / XLSX / PDF parsing
  rag.js           Builds the context/prompt from the stored statement
public/            Frontend (index.html, app.js, style.css)