Phase 5 — Reconciliation

Goal

Suggest bank matches without scanning every statement line, and refuse to guess when two candidates look equal.

Worked example (RRF arithmetic)

Expense: Hotel Adlon / 612.40 EUR / 2026-09-17

Prefilter keeps two bank rows. Ranks (k=60):

Candidate	Fuzzy rank	Amount rank	RRF score
A HOTEL ADLON…	1	1	1/61 + 1/61 ≈ 0.0328
B LUFTHANSA…	2	2	1/62 + 1/62 ≈ 0.0323

A wins. If the top-two RRF scores are within the ambiguity margin, the expense goes to pending_review instead of auto-match.

Confirm uses SELECT … FOR UPDATE so two reviewers cannot claim the same bank line. Currency mismatches are dropped in the prefilter.

Failure case

Second confirm on an already-matched expense → 409 Conflict.

RapidFuzz scores (0–100) and embedding cosine (−1–1) are not on one scale, so RRF merges ranks instead of weighted averages. Missing embeddings then degrade gracefully.

Paths: services/reconciliation.py, core/embeddings.py