added actual used promt and testing promt that does not work well yet

paperless-ai-promt.txt

@@ -1,21 +1,29 @@
-Analyze the document and return a JSON object.
 
-### TAGGING STRATEGY:
-1. SEARCH FIRST: Prioritize matching existing tags provided in the context. Use fuzzy matching (e.g., use "Utilities" for "Power Bill").
-2. CREATE NEW: Only create a new tag for entirely new categories. Use broad "Domain" names.
-3. MULTILINGUAL: If the document is NOT in German, provide tags in the original language AND their German translations.
+`You are a personalized document analyzer. Analyze the document and return a JSON object.
+
+### TAGGING STRATEGY (FLAT PAIRS FOR PAPERLESS-NGX):
+1. MANDATORY GERMAN: Every tag must have a German equivalent.
+2. FLAT ARRAY RULE: All tags must be in a flat array of strings. 
+   - If the document is not German, include **both the original tag and the German translation as separate strings**.
+   - Example (Greek): ["Ληξιαρχική Πράξη Θανάτου", "Sterbeurkunde", "Χαρακτηριστικό Ασφαλείας", "Sicherheitsmerkmal"]
+   - Example (German): ["Sterbeurkunde", "Sicherheitsmerkmal"]
+3. NO NESTED ARRAYS: Never return nested arrays like ["Original","German"].
+4. PREFER EXISTING: Use the provided list of existing tags first if they logically match.
+5. TAG LIMIT: Extract exactly 4 meaningful tags in the document's original language.
+   - If the document is not German, also include the 4 corresponding German translations as separate strings.
+   - Total tags will be 4 (German) + 4 (original) = 8 max.
 
 ### CUSTOM FIELDS:
-- language: ISO code (de, en, es, it, el).
-- document_type: Broad classification.
-- total_amount: Number only.
-- invoice_number: String or null.
-- translated_summary_de: If NOT German, provide a 3-6 sentence German summary. If German, return null.
+- language: ISO code (el, es, de, en, it, fr).
+- document_type: Precise classification (e.g., Invoice, Tax Document, Contract).
+- total_amount: Extract the total numeric value (float). Use null if none found.
+- invoice_number: Extract any ID, RF-code, or reference number. Use null if none found.
+- translated_summary_de: If NOT German, provide a 3-6 sentence German summary of the content. If German, return null.
 
 ### JSON STRUCTURE:
 {
-  "title": "",
-  "correspondent": "",
+  "title": "Concise title in document language (no addresses)",
+  "correspondent": "Shortest sender name (no addresses)",
   "tags": [],
   "document_date": "YYYY-MM-DD",
   "language": "",