Account suggestions¶

When you import a statement, Paisa can suggest an account for each transaction. Suggestions help with the first pass; you review them in the preview before anything is added to the journal.

Statement import with high, medium, unknown, and possible-transfer account suggestions

Example¶

Suppose a statement contains SWIGGY INSTAMART BANGALORE for ₹1,248. A saved merchant rule or several matching transactions in your journal may produce:

Suggested account: Expenses:Groceries
Confidence: High
Reason: known merchant with matching ledger history

An ambiguous description such as AMAZON may have been recorded under Expenses:Shopping four times and Expenses:Books three times. Paisa marks a split history like this for review instead of pretending that one answer is certain.

How a suggestion is made¶

Paisa first checks saved merchant rules, then compares the transaction with your committed ledger history. Historical matches consider the merchant, source account, direction, currency, amount, and recency. If there is still no confident match, Paisa can use an older keyword-similarity fallback.

When the evidence is weak, Paisa marks the suggestion for review or returns Unknown instead of treating it as certain.

The result includes a confidence level and a reason where available. Low- confidence and unknown suggestions should be reviewed or replaced before you save the import.

Merchant rules¶

If a merchant always belongs to the same account, save a merchant rule from the import workflow or configuration. Rules are stored under prediction.merchant_rules in paisa.yaml and take precedence over general history. Read-only configurations cannot save new rules.

Reviewing or overriding a suggestion fixes the current import. A saved merchant rule is the option that persists for future imports; Paisa does not silently change your journal categories.

Prediction runs locally on the data available to Paisa. It does not send your statement to an external AI service.

Technical notes¶

The legacy fallback represents account descriptions with TF-IDF and compares them using cosine similarity. It can produce Needs Review or Unknown, but never a high-confidence suggestion.