tocmo0nlord/odoo-ai
1
0
Fork 0
Code Issues 1 Pull Requests Actions Packages Projects Releases Wiki Activity

Fix vendor mis-identification (McDonald's bias), MIA Parking amount, grayscale OCR fallback

Browse Source 
- Remove "NeDonald's → McDonald's" from LLM vendor correction examples; the
  example was biasing the model to return McDonald's for any ambiguous receipt
  (Home Depot, Sergio's/HMSHost). Replace with neutral brand examples and add
  an explicit instruction not to substitute a brand name absent from the OCR text.
- Add `net\s*fee` to _TOTAL_RE so MIA Parking kiosk receipts ("net fee: 150.00 USD")
  are captured by Pass 1 rather than the max-scan which could pick a larger line.
- Add Step 5b grayscale fallback in receipt_parser: if all binarized PSM attempts
  yield < 20 chars, retry OCR on the pre-binarization grayscale image. Fixes
  dot-matrix and certain thermal-print fonts destroyed by the 160-threshold.
- Tests: 88 passing (test_net_fee_parking, test_vendor_prompt_does_not_contain_mcdonalds,
  test_vendor_prompt_instructs_not_to_guess_absent_brand).

Co-Authored-By: Claude Sonnet 4.6 <noreply@anthropic.com>
This commit is contained in:
Carlos Garcia
2026-05-21 00:56:45 -04:00
parent ece811cccb
commit db06fede5f
3 changed files with 109 additions and 7 deletions
Download Patch File Download Diff File Expand all files Collapse all files

20
agent_service/agents/expenses_agent.py
View File

@@ -23,7 +23,7 @@ from ..tools.expenses_tools import ExpensesTools
_TOTAL_RE = re.compile(
r'(?:grand\s*total|total\s*due|amount\s*due|balance\s*due|'
r'total\s*amount|total\s*charged|total\s*sale|net\s*sale|'
r'sale\s*total|you\s*paid|amount\s*paid|total)'
r'sale\s*total|you\s*paid|amount\s*paid|net\s*fee|total)'
r'(?!\s*tax)' # exclude "Total Tax / Total Taxes"
r'\s*[:\-]?\s*\$?\s*([\d,]+\.\d{2})',
re.IGNORECASE,
@@ -551,19 +551,25 @@ class ExpensesAgent(BaseAgent):
excerpt = stripped[:600]
prompt = (
'Return ONLY valid JSON with exactly two keys:\n'
'"vendor": the merchant or store name from the receipt header. '
'OCR often garbles text — use your knowledge to correct obvious '
'errors (e.g. "NeDonald\'s" → "McDonald\'s", "TN-N-QUT" → '
'"IN-N-OUT Burger", "Subwey" → "Subway", "LRYAL" → "LAYAL"). '
'"vendor": the business name printed at the TOP of the receipt '
'(usually the first 1-3 lines). '
'Ignore slogans ("How doers get more done"), product item names, '
'and payment-processor logos. '
'OCR often substitutes look-alike characters — correct obvious '
'errors (e.g. "LRYAL" → "LAYAL", "Subwey" → "Subway", '
'"H0ME DEP0T" → "HOME DEPOT", "W4LMART" → "WALMART"). '
'IMPORTANT: only use a brand name that is clearly present in the '
'text — do NOT substitute a different well-known brand if the '
'name is merely unclear. '
'If this looks like a bank or credit-card statement listing '
'multiple transactions rather than a single merchant receipt, '
'use "". Use "" if no clear business name is visible.\n'
f'"product_name": pick the single best match from [{product_list}]. '
'Guide: restaurant / cafe / fast food → food/meal product; '
'Guide: restaurant / cafe / fast food / food court → food/meal product; '
'airline / airport / transit / taxi / parking / rental car → travel product; '
'gas station / petrol / fuel → fuel product; '
'hotel / motel / lodging → accommodation product; '
'office / tech / hardware store → supplies product. '
'hardware / home improvement / tech / office supply store → supplies product. '
'Return "" if nothing fits.\n\n'
f'Receipt text:\n{excerpt}\n\nJSON only:'
)

18
agent_service/tools/receipt_parser.py
View File

@@ -130,6 +130,7 @@ def _ocr_image_tesseract(data: bytes, filename: str) -> str:
# ── Step 3: Grayscale + contrast ─────────────────────────────────────
img = ImageOps.grayscale(img)
img = ImageOps.autocontrast(img)
img_gray = img # save grayscale for fallback — before binarization
# ── Step 4: Sharpen then binarize ─────────────────────────────────────
# Sharpen first so edges are crisp before thresholding.
@@ -152,6 +153,23 @@ def _ocr_image_tesseract(data: bytes, filename: str) -> str:
except Exception:
pass
# ── Step 5b: Grayscale fallback ───────────────────────────────────────
# Binarization at threshold 160 can destroy dot-matrix and certain
# thermal-print fonts (e.g. parking kiosk receipts) where character
# pixels are close to the threshold and get wiped to white. If every
# binarized attempt failed, retry on the plain grayscale image —
# Tesseract handles grey-level input reasonably well for these cases.
for psm in (6, 4, 11):
try:
text = pytesseract.image_to_string(
img_gray, config=f'--oem 3 --psm {psm}').strip()
if len(text) >= 20:
logger.debug('Tesseract grayscale fallback %s: psm=%d %d chars',
filename, psm, len(text))
return text
except Exception:
pass
logger.warning('Tesseract OCR %s: all PSM modes returned < 20 chars', filename)
return ''
except ImportError:

78
tests/test_expenses_agent.py
View File

@@ -524,6 +524,18 @@ class TestExtractAmount:
text = 'Items 8.49\nCash Tendered 20.00\nChange 11.51'
assert _extract_amount_from_text(text) == 8.49
def test_net_fee_parking(self):
# Parking kiosk receipts (e.g. MIA) use "net fee: 150.00 USD" format.
# _TOTAL_RE must include "net fee" so Pass 1 catches it and avoids
# the max-scan accidentally picking up a larger line like entry/exit fees.
text = (
'MIAMI AIRPORT PARKING\n'
'Entry 05/09 08:00\n'
'Exit 05/10 14:30\n'
'net fee: 150.00 USD'
)
assert _extract_amount_from_text(text) == 150.00
class TestBankStatementDetection:
def _stmt(self, n: int) -> str:
@@ -657,6 +669,72 @@ async def test_parse_ocr_failed_skips_llm_amount():
assert result['date'] == '2026-05-10'
@pytest.mark.asyncio
async def test_vendor_prompt_does_not_contain_mcdonalds():
"""The vendor LLM prompt must not reference 'McDonald' as a correction
example — it biases the model toward returning McDonald's whenever OCR
text is unclear, causing unrelated receipts (Home Depot, HMSHost) to be
misidentified as McDonald's.
"""
agent = _make_agent()
captured: list[str] = []
llm_resp = MagicMock()
llm_resp.content = '{"vendor":"The Home Depot","product_name":"Supplies"}'
async def _capture(messages, caller=None):
for m in messages:
captured.append(m.get('content', ''))
return llm_resp
agent._llm.submit = _capture
await agent._parse_receipt_text(
'THE HOME DEPOT\nHow doers get more done\nWAGNER FURNO 300HG 36.78\nVISA USD$ 36.78',
'homedepot.jpg',
expense_products=[{'id': 1, 'name': 'Meals'}, {'id': 2, 'name': 'Supplies'}],
)
full_prompt = ' '.join(captured)
assert 'McDonald' not in full_prompt, (
"Vendor prompt must not contain 'McDonald' — it biases the model toward "
"returning McDonald's for any ambiguous receipt."
)
@pytest.mark.asyncio
async def test_vendor_prompt_instructs_not_to_guess_absent_brand():
"""Prompt must explicitly tell the LLM not to substitute a brand name that
isn't in the OCR text — prevents "default to well-known fast food" behaviour.
"""
agent = _make_agent()
captured: list[str] = []
llm_resp = MagicMock()
llm_resp.content = '{"vendor":"SERGIO\'S MIAMI AIRPORT","product_name":"Meals"}'
async def _capture(messages, caller=None):
for m in messages:
captured.append(m.get('content', ''))
return llm_resp
agent._llm.submit = _capture
await agent._parse_receipt_text(
'(((HMSHost ByAvolta\nSERGIO\'S MIAMI AIRPORT\nCHK 9745\nPayment $16.29',
'sergios.jpg',
expense_products=[{'id': 1, 'name': 'Meals'}],
)
full_prompt = ' '.join(captured)
# The prompt should warn the model not to invent brand names
assert 'only use a brand name' in full_prompt.lower() or \
'do not' in full_prompt.lower() or \
'not substitute' in full_prompt.lower(), (
"Prompt must instruct the LLM not to substitute a different brand name."
)
# ---------------------------------------------------------------------------
# parse_upload — receipt_parser.py
# ---------------------------------------------------------------------------