Document Parsing

1. Overview

Provides Claude with extraction rules for converting receipt images and PDF invoices into structured line-item data. This skill covers:

• Danish grocery receipts (thermal paper with many line items)
• Telecom invoices (PDF with service line items)
• Utility bills (PDF with consumption data)
• Restaurant receipts
• Online order confirmations

It also provides a workflow for looking up and applying vendor-specific invoice parsers from invoice-knowledge/.

---

2. Vendor Detection & Parser Lookup

Parser Lookup Workflow

When processing a receipt or invoice, follow this decision tree:

1. Detect vendor (see Vendor Detection below)
2. Normalize vendor name to vendor-id (lowercase, no spaces)
3. Check: does invoice-knowledge/{vendor-id}/PARSER.md exist?
   ├── YES: Load PARSER.md → use vendor-specific extraction rules
   │         Set confidence boost: +0.1 (parser-assisted extraction)
   └── NO:  Fall back to general rules (Section 3 and 4 below)
            Set confidence: 0.5-0.7 (general extraction)
            After processing: suggest /smartspender:receipt learn

Vendor Detection Signals

Detect the vendor using these signals in priority order:

Priority	Signal	Source	Example
1	Email sender domain	Email metadata (receipt email workflow)	faktura@tdc.dk → TDC
2	Filename	Uploaded file name	faktura_orsted_jan2026.pdf → Oersted
3	Invoice header/logo	First page content (Claude Vision)	Logo text "HOFOR" → HOFOR
4	Content keywords	Invoice body text	"TDC NET A/S" → TDC
5	Transaction match	Bank transaction description	"PBS TDC" → TDC

Vendor ID Normalization

Vendor Name	Vendor ID	Aliases
TDC	tdc	TDC NET, TDC A/S, TDC Erhverv
HOFOR	hofor	HOFOR A/S
Oersted	orsted	Oersted Salg, Oersted A/S, Oersted Danmark
Norlys	norlys	Norlys Energi, SE (former name)
Telenor	telenor	Telenor Danmark
Telia	telia	Telia Danmark, Telia Company

For vendors not in this table: normalize to lowercase, replace spaces with hyphens, remove A/S and other suffixes.

Parser Application

When a PARSER.md is found, apply it as follows:

1. Load Parser: Read invoice-knowledge/{vendor-id}/PARSER.md into context
2. Extract Metadata: Use the parser's extraction rules instead of generic rules
3. Extract Line Items: Use the parser's line items table to map invoice lines to categories
4. Apply Special Handling: Follow any special handling rules in the parser
5. Check Learned Corrections: Apply corrections from the Learned Corrections table
6. Set Confidence: Parser-assisted extractions get baseline 0.8 (vs. 0.5-0.7 for general)

Fallback Behavior

When no vendor-specific parser exists:

1. Use general extraction rules from sections 3-4 below
2. Set extraction confidence to 0.5-0.7 (lower than parser-assisted)
3. After processing completes and the user has verified/corrected the extraction, suggest:

Tip: Du kan koere /smartspender:receipt learn for at gemme
udtraeksregler for {vendor}, saa fremtidige fakturaer
fra denne leverandoer bliver mere praecise.

Only suggest this for PDF invoices (not grocery receipts or simple card slips).

Available Parsers

Vendor	Type	Parser Path
TDC	Telecom	invoice-knowledge/tdc/PARSER.md
HOFOR	Water utility	invoice-knowledge/hofor/PARSER.md
Oersted	Electricity	invoice-knowledge/orsted/PARSER.md

---

3. Grocery Receipt Extraction

Supported Receipt Types

Type	Common Sources	Key Characteristics
Grocery receipt	Foetex, Netto, Rema 1000, Bilka, Meny	Thermal paper, many line items, Danish keywords
Restaurant receipt	Restaurants, cafes	Total + tip, few items, sometimes handwritten
Online order	Amazon, Zalando, IKEA	PDF/email, shipping line, order number

Structure Recognition

Danish grocery receipts follow a consistent layout:

[Store name and address]
[Date and time]
[Cashier / register info]
---
[Item lines]
...
---
[Subtotal]
[Discounts]
[Total]
[Payment method]
[VAT summary]

Item Line Patterns

Each item line typically follows one of these formats:

Pattern	Example	Notes
Name Price	Minimælk 1L 12.95	Single item
Qty x Name Price	2 x Minimælk 1L 25.90	Multiple quantity
Qty STK Name Price	2 STK Minimælk 1L 25.90	Alternative quantity format
Name Price A/B	Minimælk 1L 12.95 A	With VAT code suffix

Danish Keywords

Keyword	Meaning	Action
I ALT	Total	Extract as receipt total
TOTAL	Total	Extract as receipt total
SUM	Sum	Extract as receipt total
RABAT	Discount	Extract as line discount
TILBUD	Offer/sale	Line has a promotional price
PANT	Bottle deposit	Extract as separate line item (category: Andet)
POSE	Bag fee	Extract as separate line item (category: Andet)
STK	Pieces (quantity)	Parse as quantity indicator
KG	Kilograms	Parse as weight-based quantity
KONTANT	Cash	Payment method — skip
DANKORT	Debit card	Payment method — skip
VISA	Credit card	Payment method — skip
MOBILEPAY	Mobile payment	Payment method — skip
MOMS	VAT	VAT summary — skip
SUBTOTAL	Subtotal	Before discounts — skip (use I ALT)
RETUR	Return	Negative line item
BYTTEPENGE	Change	Cash change — skip

Discount Handling

Discounts on Danish receipts appear in several formats:

Format	Example	Rule
Inline discount	RABAT -5.00 on next line after item	Apply discount to preceding item
Percentage	10% RABAT	Calculate discount from item price
Member discount	COOP RABAT -3.00	Apply to preceding item, note as member discount
Multi-buy	3 FOR 2 RABAT -12.95	Apply discount to the item group

Total Validation

After extracting all items, compare the sum of line totals against the receipt total:

• Variance <= 2%: Accept without comment
• Variance 2-5%: Accept but note: "Bemaerk: Summen af varerne afviger lidt fra totalen (pant, afrunding, eller manglende linje)."
• Variance > 5%: Flag for review: "Summen af varerne ({sum} kr) matcher ikke kvitteringens total ({total} kr). Tjek venligst om alle linjer er korrekt aflaest."

---

4. PDF Invoice Extraction

Structure Recognition

Danish PDF invoices typically contain:

[Company logo and name]
[Invoice number and date]
[Customer info]
[Billing period]
---
[Service line items]
---
[Subtotal before VAT]
[Moms (25%)]
[Total including moms]
[Payment info / due date]

Key Fields to Extract

Field	Where to Find	Example
Merchant	Header/logo area	TDC, Oersted, HOFOR
Invoice date	Near invoice number	Fakturadato: 15-01-2026
Billing period	Below customer info	Periode: 01-12-2025 til 31-12-2025
Line items	Main table	Mobilabonnement Frihed+ — 199,00
Moms	Before total	Moms 25% — 59,75
Total	Bottom of invoice	I alt inkl. moms — 299,00

Danish Number Formats in PDFs

Danish invoices use comma as decimal separator and period as thousands separator:

• 1.234,56 means 1234.56
• 299,00 means 299.00

Convert all amounts to standard decimal format (period as decimal separator) for storage.

---

5. Product Subcategory Taxonomy

Product-level subcategories for receipt line items. These are more granular than the merchant-level subcategories in skills/categorization/SKILL.md.

Dagligvarer (Groceries) Subcategories

Subcategory	Examples	Keywords
Mejeriprodukter	Maelk, yoghurt, ost, smoer	maelk, yoghurt, ost, smoer, floede, skyr
Koed	Hakket oksekoed, kylling, poelse	koed, kylling, svin, okse, poelse, bacon, lam
Fisk	Laks, torsk, rejer	laks, torsk, rejer, fisk, tun, sild
Frugt og groent	Bananer, aebler, tomater, salat	frugt, groent, banan, aeble, tomat, salat, loeg, kartoffel, gulerod
Broed og bagvaerk	Rugbroed, franskbroed, boller	broed, rug, bolle, kage, croissant, bagvaerk
Drikkevarer	Sodavand, juice, vand	cola, pepsi, fanta, juice, vand, sodavand, oel (non-alcoholic)
Alkohol	Vin, oel, spiritus	vin, oel, spiritus, vodka, gin, whisky, champagne
Frostvarer	Frosne pizzaer, is, frosne groentsager	frost, frossen, is, pizza (frozen)
Konserves	Daaser, pasta, ris, mel	konserves, daase, pasta, ris, mel, sukker, salt
Snacks	Chips, chokolade, slik	chips, chokolade, slik, noedder, popcorn, kiks
Personlig pleje	Shampoo, tandpasta, deodorant	shampoo, tandpasta, deodorant, saebe, creme
Rengoring	Opvaskemiddel, vaskepulver	rengoring, opvask, vaske, aftorring, toiletpapir
Husdyr	Hundefoder, kattesand	hund, kat, foder, husdyr

Bolig (Housing) Subcategories

Subcategory	Examples	Keywords
El	Elforbrug, elafregning	el, kwh, forbrug, afregning
Vand	Vandforbrug	vand, m3, forbrug
Varme	Fjernvarme, gasforbrug	varme, fjernvarme, gas
Internet	Bredbaand	internet, bredbaand, fiber, wifi

Abonnementer (Subscriptions) Subcategories

Subcategory	Examples	Keywords
Mobilabonnement	Mobiltjeneste, datatilaeg	mobil, data, tale, sms, opkald
Forsikring	Indboforsikring, mobilforsikring	forsikring, daekning
Streaming	Netflix, Spotify	(matched by merchant, not keywords)
Software	Adobe, Microsoft 365	(matched by merchant, not keywords)

---

6. Confidence Scoring

Rate the quality of each extraction:

Confidence	Criteria
0.9-1.0	Clear image/PDF, all fields extracted, total validates
0.7-0.8	Mostly readable, 1-2 items uncertain, total within 5%
0.5-0.6	Partially readable, multiple items uncertain or missing
0.3-0.4	Poor quality, only merchant/date/total extractable
0.0-0.2	Unreadable or not a receipt

Store the confidence in match_confidence on the receipts.csv row (this is extraction confidence, separate from the transaction match confidence in skills/transaction-matching/SKILL.md).

---

7. Edge Cases

Blurry or Low-Quality Receipts

If the receipt image is too blurry to extract reliably:

1. Extract whatever is readable (merchant, date, total at minimum)
2. Set confidence to 0.3-0.4
3. Set item_count to 0 if no items are readable
4. Inform user: "Kvitteringen er svaer at aflaese. Jeg kunne kun finde butik, dato og total. Vil du tilfoeje varelinjer manuelt?"

Multiple Receipts in One Image

If the image contains more than one receipt:

1. Process only the first/most prominent receipt
2. Inform user: "Billedet ser ud til at indeholde flere kvitteringer. Jeg har behandlet den foerste. Del venligst de andre enkeltvis."

Non-Danish Receipts

If the receipt is not in Danish:

1. Still attempt extraction (most receipt formats are similar)
2. Use the receipt's currency if not DKK
3. Note in file_reference: "non-danish"
4. Categories still use Danish names

Receipts Without Line Items

Some receipts only show a total (e.g., card terminal slips):

1. Create the receipts.csv row with item_count = 0
2. Skip writing to receipt-items files (don't create an empty monthly file)
3. Inform user: "Denne kvittering har kun en total — ingen varelinjer fundet."

Negative Items (Returns)

Items marked with RETUR or negative amounts:

1. Store with negative total_price
2. Category matches the original item if identifiable
3. Discount field stays 0 (the negative price is the return, not a discount)

---

8. Examples

Example 1: Grocery Receipt Extraction

Input: Photo of Foetex receipt

Extracted metadata:

merchant: Foetex
date: 2026-01-28
total: 347.50
currency: DKK
source: upload

Extracted items:

1. Minimælk 1L          qty:2  unit:12.95  total:25.90  cat:Dagligvarer  sub:Mejeriprodukter
2. Hakket oksekød 500g  qty:1  unit:45.00  total:45.00  cat:Dagligvarer  sub:Kød
3. Øko bananer          qty:1  unit:22.95  total:22.95  cat:Dagligvarer  sub:Frugt og grønt
4. Coca-Cola 1.5L       qty:2  unit:22.00  total:44.00  cat:Dagligvarer  sub:Drikkevarer
5. Rødvin Chianti       qty:1  unit:79.95  total:79.95  cat:Dagligvarer  sub:Alkohol
6. Tandbørstehoveder    qty:1  unit:89.95  total:89.95  cat:Sundhed      sub:Personlig pleje
7. Rugbrød              qty:1  unit:24.95  total:24.95  cat:Dagligvarer  sub:Brød og bagværk
8. Pose 2 kr            qty:1  unit:2.00   total:2.00   cat:Andet        sub:Andet

Sum: 334.70 vs total 347.50 — variance 3.7% — accept with note about pant/rounding.

Example 2: TDC Invoice Extraction

Input: PDF invoice from TDC

Extracted metadata:

merchant: TDC
date: 2026-01-15
total: 299.00
currency: DKK
source: upload

Extracted items:

1. Mobilabonnement Frihed+  qty:1  unit:199.00  total:199.00  cat:Abonnementer  sub:Mobilabonnement
2. Ekstra data 10GB         qty:1  unit:49.00   total:49.00   cat:Abonnementer  sub:Mobilabonnement
3. Forsikring Mobil+        qty:1  unit:51.00   total:51.00   cat:Abonnementer  sub:Forsikring

Sum: 299.00 vs total 299.00 — exact match.

Example 3: Known Vendor (TDC) with Parser

Input: PDF invoice uploaded, filename "TDC_faktura_jan2026.pdf"

Vendor detection:

1. Filename contains "TDC" → vendor candidate: TDC
2. Check invoice-knowledge/tdc/PARSER.md → exists
3. Load TDC parser

Parser-assisted extraction:

• date: Look for "Fakturadato" → 15-01-2026
• total: Look for "I alt inkl. moms" → 299,00
• Line items: Match "Mobilabonnement" → Abonnementer > Mobilabonnement

Confidence: 0.95 (all fields extracted cleanly with parser)

Example 4: Unknown Vendor (Norlys, No Parser Yet)

Input: PDF invoice from Norlys

Vendor detection:

1. Invoice header shows "Norlys" → vendor candidate: Norlys
2. Normalize to vendor-id: norlys
3. Check invoice-knowledge/norlys/PARSER.md → does not exist

Fallback extraction:

• Use general PDF invoice rules
• Extract merchant, date, total, line items using generic patterns

Confidence: 0.6 (general extraction, no vendor-specific rules)

Post-processing: After user verifies, suggest: "Tip: Du kan koere /smartspender:receipt learn for at gemme udtraeksregler for Norlys..."

---

Related Skills

• See skills/data-schemas/SKILL.md for the CSV file structure
• See skills/transaction-matching/SKILL.md for how extracted receipts are matched to bank transactions
• See skills/categorization/SKILL.md for merchant-level categories (receipt subcategories are more granular)
• See skills/email-receipt-scanning/SKILL.md for email-specific vendor detection signals