Skip to content

Magic Import (PDF-to-AAS Extraction)

Upload a PDF datasheet or nameplate and let an LLM extract field values directly into your IDTA submodel form. Magic Import provides source highlighting, confidence scoring, and full transparency over extracted data.

Magic Import Review

Key Capabilities

  • Privacy-First Extraction: Only relevant snippets are sent to the LLM, not the full document
  • Multi-Provider LLM Support: OpenAI (GPT-4o), Anthropic (Claude), OpenRouter (100+ models), or local Ollama models
  • OCR Support: Tesseract-based OCR for scanned PDFs with configurable language and DPI
  • Confidence Scoring: 4-signal weighted formula (LLM confidence, evidence match, OCR quality, format rules)
  • Source Highlighting: Click any extracted field to highlight the exact source region in the PDF viewer
  • Review Workflow: Fields below 80% confidence are flagged for human review before applying

How It Works

Upload PDF → Index Text → OCR (if needed) → Schema Resolution → BM25 Retrieval → LLM Extraction → Evidence Localization → Confidence Scoring → Review & Apply
  1. PDF Indexer: Extracts text with word-level bounding boxes using PyMuPDF
  2. OCR Engine: Falls back to Tesseract for scanned pages
  3. Schema Resolver: Enumerates target fields from the selected template with semantic hints
  4. BM25 Retriever: Finds relevant snippets using keyword matching (privacy-preserving)
  5. LLM Extractor: Extracts structured values from snippets with evidence quotes
  6. Evidence Localizer: Maps LLM quotes back to PDF coordinates using fuzzy matching
  7. Confidence Scorer: Combines signals into a single 0–1 confidence score

Confidence Scoring Formula

confidence = 0.35 × llm + 0.40 × localizer + 0.15 × ocr + 0.10 × rules
Signal Weight Description
llm 35% LLM self-reported confidence
localizer 40% Evidence quote match quality (fuzzy string matching)
ocr 15% Text extraction quality (1.0 for native PDF, lower for OCR)
rules 10% Format/type validation (dates, numbers, enums)

UI Overview

Screenshot Description
Upload Drag-drop zone for PDF upload and template selection
Review Split view: PDF viewer (left) + extraction table (right) with filter tabs
Highlight Click field → PDF evidence highlighted with quote display
Edit Inline editing mode with text input
Apply "Apply X Fields to Form" button enabled, ready to apply

Confidence Badges

The extraction table displays confidence badges indicating extraction quality:

Badge Condition Description
Edited User modified User changed the extracted value (blue)
Approved User confirmed User approved a low-confidence extraction (green)
High ≥90% confidence Auto-approved, high extraction confidence (green)
Medium 80–89% confidence Auto-approved, moderate confidence (neutral)
Low <80% confidence Needs review, amber badge with "Approve" button

Users can filter the table using tabs: All / Needs Review / Ready. The "Approve All" button bulk-approves all low-confidence fields.

Background Processing

Magic Import uses Celery + Redis for scalable job processing:

# Start with Magic Import profile
docker compose --profile magic-import up

Jobs progress through states: UPLOADED → INDEXING → OCR → EXTRACTING → LOCALIZING → SCORING → DONE

Configuration

Variable Description Default
MAGIC_IMPORT_ENABLED Enable Magic Import feature true
MAGIC_IMPORT_LLM_PROVIDER LLM provider (openai, anthropic, openrouter, local) openai
MAGIC_IMPORT_LLM_MODEL Model name gpt-4o-mini
OPENAI_API_KEY OpenAI API key -
ANTHROPIC_API_KEY Anthropic API key -
OPENROUTER_API_KEY OpenRouter API key -
OLLAMA_BASE_URL Ollama server URL http://localhost:11434
SETTINGS_STORAGE_DIR Directory for encrypted settings ./cache/settings
MAGIC_IMPORT_CONFIDENCE_THRESHOLD Flag fields below this score 0.80
MAGIC_IMPORT_OCR_ENABLED Enable OCR fallback true
MAGIC_IMPORT_OCR_LANGUAGE Tesseract language codes eng+deu
MAGIC_IMPORT_OCR_DPI OCR resolution 300
MAGIC_IMPORT_MAX_PDF_SIZE_MB Maximum PDF file size 50
MAGIC_IMPORT_JOB_TTL_HOURS Job retention period 24
CELERY_BROKER_URL Redis URL for Celery broker redis://localhost:6379/0
CELERY_RESULT_BACKEND Redis URL for Celery results redis://localhost:6379/0

LLM Provider Setup

OpenAI (default)

export MAGIC_IMPORT_LLM_PROVIDER=openai
export MAGIC_IMPORT_LLM_MODEL=gpt-4o-mini  # or gpt-4o
export OPENAI_API_KEY=sk-...

Anthropic

export MAGIC_IMPORT_LLM_PROVIDER=anthropic
export MAGIC_IMPORT_LLM_MODEL=claude-3-haiku-20240307  # or claude-3-5-sonnet-20241022
export ANTHROPIC_API_KEY=sk-ant-...

Local (Ollama)

export MAGIC_IMPORT_LLM_PROVIDER=local
export MAGIC_IMPORT_LLM_MODEL=llama3
export OLLAMA_BASE_URL=http://localhost:11434

OpenRouter

export MAGIC_IMPORT_LLM_PROVIDER=openrouter
export MAGIC_IMPORT_LLM_MODEL=anthropic/claude-3.5-sonnet  # or openai/gpt-4o, google/gemini-pro
export OPENROUTER_API_KEY=sk-or-...

LLM Settings UI

Configure LLM providers directly in the application without editing environment files.

Accessing Settings

  1. Navigate to Magic Import (Step 4 in wizard)
  2. Click "Configure Now →" in the provider status bar
  3. The LLM Provider Configuration panel opens

Provider Selection

Four providers are supported:

Provider Description API Key Required
OpenAI GPT-4o, GPT-4o-mini, GPT-4 Turbo Yes
Anthropic Claude 3.5 Sonnet, Claude 3.5 Haiku, Claude 3 Opus Yes
OpenRouter 100+ models via unified API (Claude, GPT-4, Gemini, Llama) Yes
Local (Ollama) Self-hosted models (Llama, Mistral, Mixtral) No

Configuration Workflow

  1. Select Provider — Click a provider card
  2. Enter API Key — Paste your key (masked for security)
  3. Validate — Click "Validate" to test the connection
  4. Save — Click "Save" to store encrypted credentials
  5. Select Model — Choose from provider-specific models

Security

  • API keys are encrypted at rest using Fernet (AES-128-CBC)
  • Keys are validated before storage
  • Keys are displayed masked (e.g., sk-...abc)
  • Keys are never logged or exported

Advanced Settings

Expand "Advanced Settings" to configure:

  • Confidence Threshold (50-100%) — Minimum confidence for auto-accept
  • Enable OCR — Extract text from scanned documents

Status Indicators

Status Indicator Meaning
Connected 🟢 Provider configured and healthy
Checking 🟡 Validating connection
Error 🔴 Connection failed
Not configured ⚪ No API key set

API Endpoints

  • POST /api/magic-import/jobs - Create extraction job from PDF upload
  • GET /api/magic-import/jobs/{job_id} - Get job status and progress
  • GET /api/magic-import/jobs/{job_id}/result - Get extraction results with confidence scores
  • GET /api/magic-import/jobs/{job_id}/pdf - Download PDF for viewer
  • DELETE /api/magic-import/jobs/{job_id} - Clean up job and associated files
  • GET /api/magic-import/jobs - List recent jobs
  • POST /api/magic-import/health - Service health check (LLM provider, OCR, Redis)