- Add resolveAnalysisMode() to modes/index.ts for dynamic skill lookup
- Extend structured-analysis.ts to use resolveAnalysisMode (supports custom skill prompts and skillId persistence)
- Update structured analyses route to use resolveAnalysisMode instead of hardcoded VALID_MODES set
- Add skills:read/create/edit/delete RBAC permissions to rbac.ts
- Add docs/API_GUIDE.md with full endpoint reference
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The encrypt() call threw an unhandled error when ENCRYPTION_KEY env var was missing,
causing a 500 that the frontend displayed as "Netzwerkfehler beim Speichern des Schlüssels".
Now returns a clear error message. Also fixed provider type cast that excluded 'openrouter'.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The CROSS JOIN referenced s.system_prompt which didn't exist in the
VALUES alias, causing the entire migration transaction to fail.
This prevented the skills table from being created at all.
Seed is now handled entirely by migration 0006.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
All skills API routes were using `db` directly instead of `withTenantDb`,
causing RLS to block all operations since `app.tenant_id` was never set.
This caused "Netzwerkfehler" when creating/reading skills.
Also fixes the broken seed migration (0005) which referenced a non-existent
column in the CROSS JOIN, preventing default system skills from being inserted.
New migration 0006 properly seeds the 4 default skills with full system prompts.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Replace hardcoded ANALYSIS_MODES lookups with database-driven skill loading:
- Add skills table to Drizzle schema with tenant-scoped, configurable skills
- Add analyses.skill_id FK and structured_result JSONB column
- Refactor runAnalysis()/runAnalysisSync() to resolve skills from DB
- Support skillId, skillSlug, or legacy mode enum (with fallback)
- Add structured data output via generateObject() + jsonSchema() for
skills with output_type = structured_data
- Update /api/analyses POST to accept skillId/skillSlug alongside mode
- Migration 0005: creates skills table, seeds system skills, backfills
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Add GET /api/skills read-only endpoint for fetching active tenant skills
- Update analyse-form.tsx to fetch skills dynamically, show description
as helper text, and render structured data results as table
- Extract SkillCards client component for the skill info cards
- Send skillId alongside mode slug for forward compatibility
Depends on: AIIA-97 (skills schema/API) and backend analysis integration.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Added version.json to track base version and build number
- Updated Gitea Actions workflow to bump build number on each deploy
- Footer now shows version number alongside commit hash
- Version passed as build arg through Docker build pipeline
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Integrate OpenRouter via its OpenAI-compatible API so users can select
and use OpenRouter models alongside existing Anthropic/OpenAI/Ollama
providers. Adds provider to type system, DB enum, API validation,
buildModel switch, and settings UI.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Add document download, view, and text export endpoints via ?action= query param
- Add /dokumente archive page with category filtering, pagination, and text viewer modal
- Add /analyse/[id] detail page for viewing analysis results with text export
- Make analysis list items clickable (link to detail view)
- Add download/export buttons to existing document upload component
- Add Dokumente nav item to sidebar
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Uploaded documents were lost on every redeploy because /app/uploads
had no persistent volume. Add uploads_data volume to docker-compose.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
pdfjs-dist v5 requires DOMMatrix even in legacy build. Add minimal
polyfills so PDF text extraction works in the Node.js Docker container
without @napi-rs/canvas.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The migration file was added by feat/aiia-66-source-selection but was not
registered in _journal.json, so it never runs on deploy. This caused
'source_scope' column-missing errors on document insert.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
pdf-parse v2 depends on @napi-rs/canvas (native module) which fails in
Next.js standalone Docker builds — native binaries aren't traced/copied
to the standalone output, causing DOMMatrix is not defined at runtime.
Replaced pdf-parse entirely with pdfjs-dist legacy build which works
natively in Node.js without canvas or DOM API dependencies:
- New src/lib/pdf.ts: extractTextFromPdf() using pdfjs-dist/legacy/build
- Worker file explicitly imported so Next.js file tracer includes it
- Updated all call sites: documents, norms/parse, contracts
- Removed pdf-parse from dependencies, added pdfjs-dist directly
- Changed serverExternalPackages from pdf-parse to pdfjs-dist
Verified: build succeeds, both pdf.mjs and pdf.worker.mjs present in
.next/standalone, text extraction works in standalone context.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Add DELETE /api/documents/:id endpoint that removes the DB record,
cleans up the stored file from disk, and logs an audit event. Add a
"Loeschen" button to the DokumentUpload component with confirmation
dialog.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
git is not available inside the node:20-alpine Docker image, so
git rev-parse in next.config.ts falls back to 'dev'. Now the commit
hash is passed as a COMMIT_HASH build arg from the host where git
is available, ensuring the footer displays the real commit hash.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The build hash was only in the sidebar which could be hidden or cut off.
Added a proper footer to the main content area so the commit hash is
always visible at the bottom of every dashboard page.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The old v1 API (`pdfParse(buffer)`) triggered DOMMatrix dependency via
pdfjs-dist canvas rendering path. The v2 API (`new PDFParse({ data })` +
`getText()`) uses a text-only code path that works in Node.js without
DOM/canvas polyfills.
Updated all three call sites:
- src/lib/documents/index.ts (generic document extraction)
- src/app/api/norms/parse/route.ts (norm PDF parsing)
- src/lib/contracts/index.ts (contract text extraction)
- src/types/pdf-parse.d.ts (updated type declarations for v2)
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Add DELETE /api/documents/:id endpoint that removes the DB record,
cleans up the stored file from disk, and logs an audit event. Add a
"Loeschen" button to the DokumentUpload component with confirmation
dialog.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Fix PDF extraction: detect scanned documents (no text layer), encrypted PDFs,
empty files, and missing files with clear German error messages
- Add error logging to extraction pipeline (was silently swallowed)
- Return errorMessage in document list API so UI can display failure reasons
- Add GET /api/documents/[id] endpoint for status polling
- Rewrite DokumentUpload component with:
- Auto-polling every 2s while documents are processing
- Visual step-by-step progress indicator (Hochgeladen → Extrahiere Text → Fertig)
- Error message display when extraction fails
- Debug toggle showing document ID, MIME type, size, timestamps
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Add source scope (case/global) to documents, enabling users to select
which uploaded documents the AI considers during analysis. Includes
schema migration, API support, reusable source selection UI component,
and integration into the analyse form.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
drizzle-orm v0.45.2 declares a migrator export but does not ship the
actual .js file, causing all migrations to silently fail on every deploy.
This was the root cause of the missing "documents" table (AIIA-62).
The new script reads the drizzle journal and executes SQL files directly
via pg, with its own tracking table for idempotent re-runs.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Document upload failed with "invalid input syntax for type uuid" when
decision_id or norm_instrument_id were sent as empty strings from
FormData. The ?? operator only catches null/undefined, not "".
Added emptyToUndefined() to sanitize all optional UUID fields at the
API boundary before they reach the DB insert.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Dockerfile: create /app/uploads with correct ownership before switching
to the non-root nextjs user, fixing EACCES: permission denied on mkdir
- analyse-form.tsx: send 'title' and 'query' fields (not 'question') to
match the /api/analyses endpoint contract, fixing 400 rejection that
showed "Analyse konnte nicht gestartet werden"
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Next.js 16 deprecated the "middleware" file convention in favor of "proxy".
This fixes the build warning reported during manual deploy.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
- Inject git commit short hash at build time via NEXT_PUBLIC_BUILD_HASH
- Display build hash in sidebar footer for version tracking
- Add Gitea Actions workflow to auto-deploy on push to master (SSH → pull → rebuild)
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Texts >10k chars are now split at § boundaries and parsed in separate
AI calls, then merged. This prevents maxOutputTokens truncation that
caused "AI returned invalid JSON" on large imports (~50k+ chars).
Co-Authored-By: Paperclip <noreply@paperclip.ing>
AI routes now use requirePermission() + ctx.tenantId to get the tenant,
ensuring getModelForTenant() is always called with the correct tenant ID
so that stored API keys are used. Fixes norms/parse (was falling back to
getModel()) and analyses/structured (was trusting x-tenant-id header).
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The entrypoint.sh migration script (migrate.mjs) imports
drizzle-orm/node-postgres/migrator which is never imported by app code.
Next.js standalone trace therefore omits it from node_modules, causing
the container to crash on startup before server.js runs.
Adding drizzle-orm and pg to serverExternalPackages ensures the full
packages are copied into the standalone output, fixing the migration
crash.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The "Netzwerkfehler beim Speichern des Schlüssels" was caused by two issues:
1. ENCRYPTION_KEY env var was not passed to the Docker container, so
AES-256-GCM encrypt() threw at runtime on every POST/PATCH.
2. The 0003_tenant_api_keys migration was not in the drizzle journal
and no migration runner existed in the Docker image.
Changes:
- docker-compose.yml: pass ENCRYPTION_KEY to app container
- .env.example: document ENCRYPTION_KEY with generation command
- .gitignore: allow .env.example to be tracked
- Dockerfile: include drizzle/ migrations and entrypoint script
- entrypoint.sh: run migrations before starting the app
- migrate.mjs: runtime migration script using drizzle-orm migrator
- drizzle journal: register 0003_tenant_api_keys migration
Co-Authored-By: Paperclip <noreply@paperclip.ing>
When selecting Anthropic or OpenAI as provider, the form now shows
an inline API key input field that creates or replaces the key via
the existing CRUD API. Includes key hint display and status indicator.
Resolves AIIA-49
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Zeigt im AI-Provider Formular einen Status-Indikator an, ob ein API-Schluessel
fuer den gewaehlten Provider (Anthropic/OpenAI) hinterlegt ist. Entfernt Ollama
aus dem API-Key Dropdown, da Ollama keine API-Keys benoetigt.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
The import page now accepts PDF files in addition to TXT.
Backend uses pdf-parse to extract text from uploaded PDFs
before sending to AI for paragraph parsing.
Co-Authored-By: Paperclip <noreply@paperclip.ing>
Neuer /normen/import Bildschirm: Gesetzestext als Fliesstext einfuegen
oder TXT-Datei hochladen, KI zerlegt automatisch in Paragraphen,
Vorschau mit Bearbeitungsmoeglichkeit, dann Import ins Regelwerk.
- POST /api/norms/parse: AI-gestuetztes Parsing von Gesetzestexten
- /normen/import: Mehrstufiges Frontend (Eingabe -> Vorschau -> Import)
- Link zum Fliesstext-Import auf der Normen-Uebersichtsseite
Co-Authored-By: Paperclip <noreply@paperclip.ing>
