From 12576ead1b8bac03dfd8b6d78bb0876e937a3dd5 Mon Sep 17 00:00:00 2001
From: Carlos Garcia <tocmo@DESKTOP-O6U0UOS.AVC.local>
Date: Sat, 16 May 2026 16:48:51 -0400
Subject: [PATCH] feat: two-pass dedup catches same-vendor OCR amount misreads

Pass 1 unchanged: same date + amount within 0.05 + vendor similarity 60%.
Pass 2 (new): same vendor (>= 80% similarity) + same date, regardless
of amount, to catch receipts where OCR misread the total.

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
---
 agent_service/agents/expenses_agent.py | 60 ++++++++++++++++++--------
 1 file changed, 41 insertions(+), 19 deletions(-)

diff --git a/agent_service/agents/expenses_agent.py b/agent_service/agents/expenses_agent.py
index 09fb2d0..53756d6 100644
--- a/agent_service/agents/expenses_agent.py
+++ b/agent_service/agents/expenses_agent.py
@@ -254,12 +254,17 @@ class ExpensesAgent(BaseAgent):
         Return the index in `candidates` of a receipt that appears to be the
         same physical receipt as `parsed`, or None if no match found.
 
-        Match criteria (all must pass):
+        Pass 1 — exact-amount match (all must pass):
           1. Same date
           2. Amount > 0 and within $0.05 of each other
-          3. Transaction times within 30 min of each other (if both present);
-             times > 30 min apart rule out a duplicate
-          4. Vendor name similarity >= 60 %  (or both vendors are raw filenames)
+          3. Transaction times within 30 min (if both present)
+          4. Vendor similarity >= 60 %  (or both vendors are raw filenames)
+
+        Pass 2 — OCR-error match (amount may differ due to misread):
+          1. Same date
+          2. Both amounts > 0
+          3. Vendor similarity >= 80 % (stricter threshold compensates for loose amount)
+          4. Times within 30 min (if both present)
         """
         amt = float(parsed.get('amount', 0))
         date = parsed.get('date', '')
@@ -267,35 +272,52 @@ class ExpensesAgent(BaseAgent):
         vendor = str(parsed.get('vendor', '')).lower().strip()
         is_filename = vendor.endswith(('.jpg', '.jpeg', '.png', '.gif', '.webp', '.bmp'))
 
+        def _times_compatible(t1, t2) -> bool:
+            """Return False only when both times are present and >30 min apart."""
+            if not (t1 and t2):
+                return True
+            try:
+                h1, m1 = (int(p) for p in t1.split(':')[:2])
+                h2, m2 = (int(p) for p in t2.split(':')[:2])
+                return abs((h1 * 60 + m1) - (h2 * 60 + m2)) <= 30
+            except Exception:
+                return True
+
+        # Pass 1: amount must match within $0.05
         for idx, (_, other) in enumerate(candidates):
             other_amt = float(other.get('amount', 0))
-            # Skip zero-amount receipts — too ambiguous to dedup
             if amt == 0 or other_amt == 0:
                 continue
             if abs(amt - other_amt) > 0.05:
                 continue
             if date != other.get('date', ''):
                 continue
-            # Time check: if both receipts have a transaction time and they are
-            # more than 30 minutes apart they are different transactions.
-            other_time = other.get('time')
-            if time and other_time:
-                try:
-                    h1, m1 = (int(p) for p in time.split(':')[:2])
-                    h2, m2 = (int(p) for p in other_time.split(':')[:2])
-                    if abs((h1 * 60 + m1) - (h2 * 60 + m2)) > 30:
-                        continue
-                except Exception:
-                    pass  # unparseable time — ignore the signal
+            if not _times_compatible(time, other.get('time')):
+                continue
             other_vendor = str(other.get('vendor', '')).lower().strip()
             other_is_filename = other_vendor.endswith(
                 ('.jpg', '.jpeg', '.png', '.gif', '.webp', '.bmp'))
             if is_filename or other_is_filename:
-                # Same amount + date, no vendor text to compare — treat as dup
                 return idx
-            ratio = difflib.SequenceMatcher(None, vendor, other_vendor).ratio()
-            if ratio >= 0.6:
+            if difflib.SequenceMatcher(None, vendor, other_vendor).ratio() >= 0.6:
                 return idx
+
+        # Pass 2: same vendor + same date even when amounts differ (OCR misread)
+        if not is_filename:
+            for idx, (_, other) in enumerate(candidates):
+                other_amt = float(other.get('amount', 0))
+                if amt == 0 or other_amt == 0:
+                    continue
+                if date != other.get('date', ''):
+                    continue
+                if not _times_compatible(time, other.get('time')):
+                    continue
+                other_vendor = str(other.get('vendor', '')).lower().strip()
+                if other_vendor.endswith(('.jpg', '.jpeg', '.png', '.gif', '.webp', '.bmp')):
+                    continue
+                if difflib.SequenceMatcher(None, vendor, other_vendor).ratio() >= 0.80:
+                    return idx
+
         return None
 
     async def _parse_receipt_text(self, text: str, filename: str,