init Command¶

Overview¶

The init command creates a config.yaml file in your current directory through an interactive setup process.

Purpose:

Generate configuration files
Validate dependencies
Guide API key setup
Provide next steps

Basic Usage¶

docling-graph init

This launches an interactive wizard that guides you through:

Processing mode (one-to-one / many-to-one)
Extraction contract (direct / staged / delta)
Delta Extraction Tuning (only when delta is selected): resolvers (enable + mode) and optional quality gate tweaks
Backend type (LLM / VLM)
Inference location (local / remote; skipped for VLM)
Docling pipeline and export options
Provider and model selection (by backend/inference)
Export format
Output directory
(If remote LLM with custom provider) Use custom endpoint (URL and API key via environment variables)

When custom endpoint is enabled, the wizard expects the fixed env var names CUSTOM_LLM_BASE_URL and CUSTOM_LLM_API_KEY.

Interactive Setup¶

Step 1: Processing Mode¶

1. Processing Mode
 How should documents be processed?
  • one-to-one: Creates a separate Pydantic instance for each page
  • many-to-one: Combines the entire document into a single Pydantic instance
Select processing mode (one-to-one, many-to-one) [many-to-one]:

Step 2: Extraction Contract¶

2. Extraction Contract
 How should LLM extraction prompts/execution be orchestrated?
  • direct: Single-pass best-effort extraction (fastest)
  • staged: Multi-pass staged extraction for improved small-model quality
  • delta: Chunk-based graph extraction for long documents (batches → merge → projection)
Select extraction contract (direct, staged, delta) [direct]:

Step 3: Delta Extraction Tuning (only when delta is selected)¶

If you selected delta, the wizard shows a Delta Extraction Tuning section:

Delta Resolvers — Enable post-merge duplicate resolution and choose mode (off, fuzzy, semantic, chain). Resolvers merge near-duplicate entities after the graph merge.
Delta Quality Gates — Optionally customize:
Max parent lookup misses before failing the quality gate
Adaptive parent lookup tolerance
Require at least one relationship in the projected graph
Require structural attachments (avoid root-only outputs)

These values are written into defaults in your config.yaml. You can override them later in the config file or via the convert command (where applicable). See Delta Extraction for full options.

Step 4: Backend Type¶

Backend Type
 Which AI backend should be used?
  • llm: Language Model (text-based)
  • vlm: Vision-Language Model (image-based)
Select backend type [llm]:

Step 5: Inference Location (LLM only)¶

Inference Location
 How should models be executed?
  • local: Run on your machine
  • remote: Use cloud APIs
Select inference location [remote]:

(VLM backend skips this step and uses local inference.)

Step 6: Docling Pipeline and Export Options¶

You choose the Docling pipeline (ocr / vision) and whether to export Docling JSON, markdown, and per-page markdown.

Step 7: Provider and Model Selection¶

For Local LLM:

Choose local LLM provider:
1. vLLM (recommended for GPU)
2. Ollama (recommended for CPU)
3. LM Studio (OpenAI-compatible local server)
4. Custom

Your choice [1-4]:

For Remote LLM:

Choose remote provider:
1. Mistral AI
2. OpenAI
3. Google Gemini
4. IBM watsonx

Your choice [1-4]:

Step 8: Model Selection¶

Select model for <provider> [default]:

Step 9: Export Format¶

Export Format
 Output format for results
  • csv: CSV files (nodes.csv, edges.csv)
  • cypher: Cypher script for Neo4j
Select export format [csv]:

Step 10: Output Directory¶

The wizard then prompts for the output directory (default: outputs).

Generated Configuration¶

Example: Remote LLM (Mistral)¶

# config.yaml
defaults:
  processing_mode: many-to-one
  backend: llm
  inference: remote
  export_format: csv

docling:
  pipeline: ocr
  export:
    docling_json: true
    markdown: true
    per_page_markdown: false

models:
  llm:
    local:
      model: ibm-granite/granite-4.0-1b
      provider: vllm
    remote:
      model: mistral-small-latest
      provider: mistral
  vlm:
    local:
      model: numind/NuExtract-2.0-8B
      provider: docling

output:
  directory: outputs

Example: Local LLM (Ollama)¶

defaults:
  processing_mode: many-to-one
  backend: llm
  inference: local
  export_format: csv

models:
  llm:
    local:
      model: llama3:8b
      provider: ollama
    remote:
      model: mistral-small-latest
      provider: mistral
  vlm:
    local:
      model: numind/NuExtract-2.0-8B
      provider: docling

output:
  directory: outputs

Example: Local LLM (LM Studio)¶

defaults:
  processing_mode: many-to-one
  backend: llm
  inference: local
  export_format: csv

models:
  llm:
    local:
      model: llama-3.2-3b-instruct   # Must match model name in LM Studio
      provider: lmstudio
    remote:
      model: mistral-small-latest
      provider: mistral
  vlm:
    local:
      model: numind/NuExtract-2.0-8B
      provider: docling

output:
  directory: outputs

Example: VLM (Local)¶

defaults:
  processing_mode: one-to-one
  backend: vlm
  inference: local
  export_format: csv

docling:
  pipeline: vision

models:
  llm:
    local:
      model: ibm-granite/granite-4.0-1b
      provider: vllm
    remote:
      model: mistral-small-latest
      provider: mistral
  vlm:
    local:
      model: numind/NuExtract-2.0-8B
      provider: docling

output:
  directory: outputs

Example: Delta extraction with resolvers¶

When you select delta and enable resolvers (and optionally customize quality), your defaults will include delta options:

defaults:
  processing_mode: many-to-one
  extraction_contract: delta
  delta_resolvers_enabled: true
  delta_resolvers_mode: semantic
  delta_quality_max_parent_lookup_miss: 4
  delta_quality_adaptive_parent_lookup: true
  delta_quality_require_relationships: false
  delta_quality_require_structural_attachments: false
  backend: llm
  inference: remote
  export_format: csv
# ... docling, models, output

You can then run docling-graph convert ... --extraction-contract delta and the config will be used. See Delta Extraction and convert Command — Delta Extraction Tuning for details.

Dependency Validation¶

After configuration, init validates required dependencies:

All Dependencies Installed¶

✅ All required dependencies are installed

Missing Dependencies¶

⚠ Missing dependencies for remote inference
Run: pip install docling-graph

Dependencies: - pip install docling-graph installs the package with LiteLLM and all core runtime dependencies. If you installed from source, use uv sync instead.

Next Steps Guidance¶

Remote Provider Setup¶

Next steps:
1. Install (if not already): pip install docling-graph

2. Set your API key:
   export MISTRAL_API_KEY="your-api-key-here"

   (If you chose custom endpoint, set instead:)
   export CUSTOM_LLM_BASE_URL="https://your-llm.example.com/v1"
   export CUSTOM_LLM_API_KEY="your-key"

3. Run your first conversion:
   docling-graph convert document.pdf \
       --template "templates.BillingDocument"

Local Provider Setup¶

If you selected LM Studio: start the Local Server in the LM Studio app; set LM_STUDIO_API_KEY only if your server requires authentication.

Next steps:
1. Install (if not already): pip install docling-graph

2. Start Ollama server (if using Ollama):
   ollama serve

3. Pull the model (if using Ollama):
   ollama pull llama3:8b

4. Run your first conversion:
   docling-graph convert document.pdf \
       --template "templates.BillingDocument"

Overwriting Configuration¶

If config.yaml already exists:

A configuration file: 'config.yaml' already exists.
Overwrite it? [y/N]:

y - Replace existing configuration
N - Cancel and keep existing file

Non-Interactive Mode¶

If interactive mode is unavailable (e.g., in CI/CD):

docling-graph init
# Falls back to default configuration

Default configuration uses: - Processing: many-to-one - Extraction contract: direct - Backend: llm - Inference: local - Provider: vllm - Export: csv

Complete Examples¶

📍 First-Time Setup¶

# Install (if not already)
pip install docling-graph

# Initialize configuration
docling-graph init

# Follow prompts:
# 1. Processing mode (e.g. many-to-one)
# 2. Extraction contract (direct / staged / delta)
# 3. (If delta) Resolvers and quality tuning
# 4. Backend (e.g. LLM), inference (e.g. remote)
# 5. Docling pipeline and export options
# 6. Provider and model
# 7. Export format, output directory

# Set API key
export MISTRAL_API_KEY="your-key"

# Test conversion
docling-graph convert test.pdf \
    --template "templates.BillingDocument"

📍 Local Development Setup¶

# Install (if not already)
pip install docling-graph

# Initialize for local development
docling-graph init

# Follow prompts:
# 1. Processing mode (e.g. many-to-one)
# 2. Extraction contract (e.g. direct)
# 3. Backend (LLM), inference (local)
# 4. Docling pipeline and export options
# 5. Provider and model
# 6. Export format, output directory

# Start Ollama
ollama serve

# Pull model
ollama pull llama3:8b

# Test conversion
docling-graph convert test.pdf \
    --template "templates.BillingDocument"

📍 VLM Setup¶

# Install (if not already)
pip install docling-graph

# Initialize for VLM
docling-graph init

# Follow prompts:
# 1. Processing mode (e.g. one-to-one)
# 2. Extraction contract (e.g. direct)
# 3. Backend (VLM) — inference is local only
# 4. Docling pipeline and export options
# 5. Model, export format, output directory

# Test conversion
docling-graph convert form.jpg \
    --template "templates.IDCard"

Configuration File Location¶

The config.yaml file is created in your current working directory:

# Create config in project root
cd /path/to/project
docling-graph init

# Creates: /path/to/project/config.yaml

Best Practice: Run init from your project root directory.

Manual Configuration¶

You can also create config.yaml manually:

# Minimal configuration
defaults:
  backend: llm
  inference: remote

models:
  llm:
    remote:
      model: mistral-small-latest
      provider: mistral

Or use the template:

# Copy template
cp docling_graph/config_template.yaml config.yaml

# Edit as needed
nano config.yaml

Troubleshooting¶

🐛 Interactive Mode Not Available¶

Error:

Interactive mode not available. Using default configuration.

Solution: - Running in non-interactive environment (CI/CD) - Default configuration will be used - Manually edit config.yaml if needed

🐛 Permission Denied¶

Error:

Error saving config: Permission denied

Solution:

# Check directory permissions
ls -la

# Run from writable directory
cd ~/projects/my-project
docling-graph init

🐛 Invalid Configuration¶

Error:

Error creating config: Invalid backend type

Solution: - Restart init command - Choose valid options (llm/vlm) - Check for typos in manual edits

Best Practices¶

👍 Initialize Per Project¶

# ✅ Good - One config per project
cd project1/
docling-graph init

cd project2/
docling-graph init

# ❌ Avoid - Shared config across projects
cd ~/
docling-graph init

👍 Version Control¶

# ✅ Good - Track configuration
git add config.yaml
git commit -m "Add docling-graph configuration"

# Add to .gitignore if it contains secrets
echo "config.yaml" >> .gitignore

👍 Environment-Specific Configs¶

# Development
cp config.yaml config.dev.yaml

# Production
cp config.yaml config.prod.yaml

# Use specific config
cp config.prod.yaml config.yaml
docling-graph convert document.pdf -t "templates.BillingDocument"

Next Steps¶

Now that you have a configuration:

convert Command → - Convert documents
CLI Recipes → - Common patterns
Configuration Guide → - Advanced config