Summarizer

Generate summaries for short and long documents. Uses Generative AI to produce concise document summaries without requiring training data.

Overview

The Summarizer processor uses LLM-based Generative AI to produce concise summaries of documents. Unlike other custom processor types, the Summarizer works out of the box without training — it uses pre-configured prompt templates and language models to generate summaries immediately.

Use cases:

• Document digests for quick review
• Executive summaries of lengthy reports
• Content previews in document management systems
• Meeting notes condensation
• Legal document summarization
• Research paper abstracts

Creating a Summarizer

Navigate to Processors → Custom Processors and click Create processor on the Summarizer card. Enter a name (e.g., “Report Summarizer”) and click Create.

The Summarizer is ready to use immediately — no training data or labeling required.

How It Works

The Summarizer uses the LLM annotator pipeline in marie-ai:

1. Document ingestion — OCR text extracted from document pages
2. Context assembly — Page text aggregated with document metadata
3. Prompt rendering — Jinja2 summary template populated with document content
4. LLM generation — Language model generates summary from prompt
5. Output formatting — Summary returned in configured format (markdown or plain text)

Configuration

The Summarizer is configured through the annotator pipeline:

Supported Models

The Summarizer works with any configured LLM connection:

• DeepSeek R1 (32B)
• Qwen 2.5 VL (vision-language)
• OpenAI GPT-4
• Anthropic Claude
• Any model configured in LLM Connections

Integration with Workflows

Add the Summarizer as a workflow node to automatically generate summaries during document processing:

Document Upload → OCR → Summarizer → Store Summary
                                    → Notify Stakeholders

Common workflow patterns:

• Intake summary — Summarize uploaded documents for quick triage
• Post-extraction summary — Generate summary after field extraction for human review
• Batch processing — Summarize all documents in a submission
• Conditional summary — Only summarize documents exceeding a page threshold

Output

The Summarizer produces:

• Summary text — Concise document summary in markdown or plain text
• Confidence — Generation confidence score
• Token usage — Input and output token counts for cost tracking

Output is accessible in downstream workflow nodes via JSONPath.

Best Practices

1. Choose the right model — Larger models produce better summaries for complex documents
2. Set appropriate length — Configure max tokens based on your summary needs
3. Use in pipelines — Combine with classifiers and extractors for comprehensive processing
4. Monitor costs — Track token usage in LLM Observability
5. Review quality — Spot-check summaries to ensure accuracy for your document types

Comparison with Other Processor Types

Next Steps

• Configure LLM Connections for model access
• Build summarization Workflows
• Monitor costs in LLM Observability
• Explore Custom Extractor for structured data extraction