Configuring NumbyAI for Real Use

If you want to test a finance AI app today, NumbyAI is interesting because it is practical and honest about the messy real world. Bank CSVs are not standard. Dates come in multiple shapes. Your biggest risk is not model quality — it is inconsistent configuration.

Configuration order

Tune in this order: Date format handling → Numeric parsing → Delimiter and preamble rules → Debit/credit and balance interpretation → Currency detection. Date parsing is the first dependency for time ordering. Numeric parsing follows because a wrong decimal assumption breaks budget summaries silently.

Database decisions

Use SQLite when learning, evaluating, or the expected input size is small. Use PostgreSQL when planning multi-user workflows or caring about long-term backup patterns. Pick one strategy and hardcode it in your setup notes before import tests.

LLM setup

Install Ollama and pull the model fully before starting the app. Keep fallback reserved for ambiguous row patterns and edge case strings — not as default path for all rows. If you configure fallback as default, you lose speed and consistency.

Pre-flight checklist

Before importing the first real statement: - Can I start the stack without errors? - Is Ollama online and model loaded? - Is .env clearly reflecting chosen DB? - Do imported rows with EU commas and US dots parse correctly? - Do files with metadata preamble rows skip the header noise? - Does debit/credit split behave correctly? - Does fallback trigger only on uncertain rows?