Privacy Levels

The privacy level controls what data is sent to an AI provider when generating insights or answering questions. This is a security-critical setting — choose it carefully.

The five levels

disabled — "Nothing sent"

No data is sent to any AI provider. The AI subsystem is completely off. This is the right choice if you don't want any AI features.

What leaves your server: nothing.

generalizations_only — "Category names and patterns, no amounts, no merchants"

Abstract patterns only. The AI sees high-level behavioral summaries — "spending in category X increased this month" — with no specific amounts, no merchant names, and no dates beyond period granularity.

What leaves your server: abstract spending patterns, category names, period-level trends.
What does NOT leave: amounts, merchant names, account names, exact dates, or any identifying information.

This is the default for all cloud providers (Anthropic, OpenAI). It cannot be set lower than this for cloud providers.

aggregates_only — "Category totals and stats, no individual transactions"

Category-level spending totals and aggregate statistics. Amounts are included at the category level. Merchant names are one-way hashed (cannot be reversed by anyone, including Anthropic or OpenAI).

What leaves your server: category totals, aggregate amounts, hashed merchant names.
What does NOT leave: individual transaction amounts, raw merchant names, account numbers, income source identities.

redacted — "Transaction structure, amounts, hashed merchants"

Transaction-level data with personally identifiable information removed. Individual transaction amounts are included. Account numbers, full transaction descriptions, and income source identities are stripped.

What leaves your server: transaction amounts and dates, hashed merchant names, transaction types (payroll/grocery/dining/etc.).
What does NOT leave: account numbers, full descriptions, income source names.

full — "Everything — local providers only"

All data, unsanitized. Only available for local providers (Ollama, llama.cpp). Hard-blocked for cloud providers (Anthropic, OpenAI) regardless of user setting. Even if you set full for a cloud provider in configuration, it will never be sent — the redaction layer enforces this in code.

What leaves your server: complete transaction history, all amounts, all merchant names, all account balances.
What does NOT leave: anything (everything stays on your server with a local model).

What "hashed merchant" means

A hashed merchant name is a one-way transformation. "Target" might become a7f3bc.... The AI provider sees a consistent identifier for each merchant (useful for patterns — "you spent at [merchant-hash-x] 12 times this month") without seeing the actual name.

This hash cannot be reversed — not by the AI provider, not by us, not by anyone. It's a one-way function applied before any data leaves your server.

Choosing a level

For cloud providers (Anthropic, OpenAI): - generalizations_only is the default — safe for most use cases, gives useful high-level insights - aggregates_only allows some amount-level insights if you're comfortable with category totals leaving your server - redacted gives the AI more to work with but sends individual transaction amounts

For local providers (Ollama, llama.cpp): - full is available and recommended — data never leaves your server at any privacy level - Setting a lower level is possible if you want to limit what the local model can see

Why full is blocked for cloud providers

Sending your complete financial transaction history to a cloud API provider — even a privacy-conscious one — is a significant data exposure. WDIAG enforces this at the application layer, not just via documentation. Even if you deliberately set full for a cloud provider, the redaction layer will cap it at redacted. This is not configurable.

If you want full access to your data for AI, use a local model.