opus is dumb

modules/_hosts/tahani/paperless-gpt-prompts/tag_prompt.tmpl

@@ -0,0 +1,26 @@
+I will provide you with the content and title of a document. Your task is to select appropriate tags for the document from the available list.
+Only select tags from the provided list.
+
+Rules:
+1. Focus on WHAT the document IS (document type) and what TOPIC it relates to — not on incidental details mentioned in the content.
+   - GOOD tags for a server hosting invoice: "Invoice", "Hosting"
+   - BAD tags for a server hosting invoice: "IBAN", "VAT", "Bank account" — these are just details that appear on any invoice.
+2. Pick 1-4 tags maximum. Fewer is better. Every tag must add distinct, meaningful categorisation value.
+3. All tags must be in English.
+4. Never tag based on formatting details, payment methods, reference numbers, or boilerplate text.
+
+The content is likely in {{.Language}}, but tags must always be in English.
+
+<available_tags>
+{{.AvailableTags | join ", "}}
+</available_tags>
+
+<title>
+{{.Title}}
+</title>
+
+<content>
+{{.Content}}
+</content>
+
+Respond only with the selected tags as a comma-separated list, without any additional information.

modules/_hosts/tahani/paperless-gpt-prompts/title_prompt.tmpl

@@ -0,0 +1,26 @@
+I will provide you with the content of a document that has been partially read by OCR (so it may contain errors).
+Your task is to generate a clear, consistent document title for use in paperless-ngx.
+
+Title format: "YYYY-MM-DD - Sender - Description"
+- YYYY-MM-DD: The document date (issue date, statement date, etc.). Use the most specific date available. If no date is found, omit the date prefix.
+- Sender: The company, organisation, or person who sent/issued the document. Use their common short name (e.g. "Hetzner" not "Hetzner Online GmbH").
+- Description: A brief description of what the document is (e.g. "Server hosting invoice", "Payslip January", "Employment contract", "Tax assessment 2024"). Keep it concise but specific enough to distinguish from similar documents.
+
+Examples:
+- "2025-03-01 - Hetzner - Server hosting invoice"
+- "2024-12-15 - Techniker Krankenkasse - Health insurance statement"
+- "2024-06-30 - Acme Corp - Payslip June"
+- "2024-01-10 - Finanzamt Berlin - Tax assessment 2023"
+
+Rules:
+1. Always write the title in English, regardless of the document language.
+2. Keep the description part under 6 words.
+3. If the original title contains useful information, use it to inform your suggestion.
+4. Respond only with the title, without any additional information.
+
+The content is likely in {{.Language}}.
+
+<original_title>{{.Title}}</original_title>
+<content>
+{{.Content}}
+</content>

modules/_hosts/tahani/paperless.nix

@@ -34,6 +34,8 @@
 			volumes = [
 				"paperless-gpt-data:/app/data"
 				"paperless-gpt-prompts:/app/prompts"
+				"${./paperless-gpt-prompts/tag_prompt.tmpl}:/app/prompts/tag_prompt.tmpl:ro"
+				"${./paperless-gpt-prompts/title_prompt.tmpl}:/app/prompts/title_prompt.tmpl:ro"
 			];
 			environment = {
 				PAPERLESS_BASE_URL = "http://host.docker.internal:${toString config.services.paperless.port}";