Complex data extraction and orchestration framework designed for processing unstructured documents. It integrates AI-powered document pipelines (GenAI, LLM, VLLM) into your applications, supporting various tasks such as document cleanup, optical character recognition (OCR), classification, splitting, named entity recognition, and form processing
Marie-AI is an Agentic Document Intelligence Platform. It orchestrates a network of autonomous, specialized AI agents—each responsible for tasks like OCR, classification, NER, and document transformation—delivering robust, scalable, and extensible document understanding for modern enterprise needs.
Autonomous Agents: Modular executors act as agents, each with clear responsibilities.
Composable Workflows: Agents are orchestrated and dynamically routed based on your goals.
Extensible & Modular: Easily add, swap, or customize agents for new workflows.
Goal-Driven Automation: Instruct Marie-AI with high-level objectives and let agents handle the execution.
Marie-AI brings the next generation of agentic AI to document processing and intelligence!
🚀 Features
Agent-based Architecture: Modular executors/agents for OCR, classification, NER, transformation, and more.
End-to-End Automation: From ingestion to extraction and delivery, with minimal human intervention.
Job Scheduling & SLAs: Advanced job orchestration with support for soft and hard Service Level Agreements.
DAG-Driven Execution: Workflows and jobs are managed as Directed Acyclic Graphs (DAGs) for dependency-aware, parallel, and fault-tolerant execution.
Runtime Fabric Operations: Analyze gateways, executors, scheduler pressure, jobs, DAGs, throughput, and dependency health with the read-only m3top terminal.
Flexible Deployment: Use as a Python package, Docker container, or standalone service.
CLI Utility:marie command-line tool for streamlined local or remote processing.
Production Hardened: Robust release/versioning, CI/CD, observability, and telemetry.
🏥 Real-World Usage: Document Extraction from Unstructured Data
Marie-AI excels at extracting structured information from unstructured or semi-structured documents across diverse industries, including:
Healthcare: Extract patient data, clinical findings, and insurance details from forms, medical records, and lab reports.
Legal: Parse contracts, agreements, and case documents for entity extraction, clause identification, and compliance checks.
Real Estate: Digitize and extract fields from deeds, lease agreements, mortgage forms, and appraisals.
Finance: Automate extraction from invoices, receipts, KYC documents, and bank statements.
Government/Education: Process applications, certificates, and official records at scale.
Whether your data lives in scans, PDFs, or complex document batches, Marie-AI’s agentic pipeline automates the end-to-end process: OCR, classification, entity extraction, table/form detection, and custom business logic.
🔗 Model Support & LLM Integration
Marie-AI works seamlessly with a variety of language models:
Off-the-shelf LLMs: Instantly connect to state-of-the-art models such as GPT, T5, BERT, and more for fast setup and broad generalization.
Fine-tuned/Custom LLMs: Easily integrate your own or third-party fine-tuned models to enable domain-specific extraction, reasoning, and compliance—perfect for specialized use cases in healthcare, legal, real estate, and beyond.
Plug-and-play architecture: Swap models or orchestrate multiple agents powered by different LLMs within a single workflow, ensuring flexibility as your needs and the LLM landscape evolve.
With Marie-AI, your document intelligence pipeline can adapt to the latest advances in large language models—whether you need generic understanding or highly specialized, compliant extraction.
⏱️ Job Scheduling, SLAs, and DAGs
Marie-AI includes a production-grade job scheduling subsystem for managing and orchestrating document processing workflows at scale.
Job Scheduler: Schedule, enqueue, and monitor jobs with support for immediate or delayed execution.
SLA Management: Define soft and hard SLAs per job or workflow, ensuring deadline-driven execution and automatic escalation or retries.
DAG Workflows: Jobs are executed as Directed Acyclic Graphs (DAGs), allowing complex dependency management, parallel step execution, and robust error handling. Each node in the DAG can represent an agent or processing step.
Policies & Retry: Fine-grained control over job policies, priority, backoff strategies, and retries.
Monitoring: Real-time state tracking for all jobs and DAGs, with SQL and API interfaces for querying job status, logs, and history.
Example:
Submit a batch extraction job as a DAG. The scheduler will coordinate all dependent steps, track their state, and enforce SLA requirements.
Use SQL or API to monitor progress and handle maintenance (pause, stop, restart, or purge jobs).
🖥️ Runtime Fabric Operations with m3top
m3top is Marie-AI's standalone operational terminal for Runtime Fabric. It connects directly to a Marie gateway and presents live executor capacity, scheduler state, job and DAG lifecycles, worker execution history, completion throughput, and dependency health without requiring M3 Forge or a browser.
The walkthrough cycles through four live operational views: executor fleet capacity, selected executor detail, DAGs requiring attention, and retained worker execution history.
Use it to distinguish capacity saturation from missing executors, investigate stalled or failed work, inspect retained execution attempts, and verify whether scheduler activity is producing durable completions. Missing measurements remain N/A; m3top does not turn unavailable telemetry into synthetic zeroes.
Static Linux binaries are published for x86-64 and ARM64 with SHA-256 checksums. From a Marie-AI checkout, install the latest m3top-v* release with:
./scripts/install-m3top.sh
m3top --version
See the m3top Operational Terminal guide for direct release installation, gateway configuration, keyboard controls, metric semantics, security boundaries, and troubleshooting.
🛠️ Skills
Marie-AI includes a skill system for packaging domain expertise as reusable, discoverable capabilities. Skills follow the ADK specification and use progressive disclosure to minimize token usage.
Agent-facing tools:
| Tool | Purpose | Token Cost |
|------|---------|------------|
| discover_skills | Search available skills by name, tags, or query | ~100 tokens/skill |
| load_skill | Load full instructions for a specific skill | <5,000 tokens |
| load_skill_resource | Load reference docs, scripts, or templates | Varies |
docker run --rm -it --entrypoint /bin/bash marieai/marie:3.0.30-cuda
marie server --start --uses sample.yml
🖥️ Command-Line Interface
Interact with the API from your terminal:
marie -h
🔥 Example Use Cases
Autonomous document cleanup and OCR
Agent-driven document classification and splitting
Named Entity Recognition (NER) as an agent
Form and table detection
Custom agent pipeline composition
See MarieAI docs for full code samples and advanced scenarios.
🔄 Release & Versioning
Marie is released via PyPI and Docker Hub. See RELEASE.md for details on versioning, release cycles, and manual release workflows.
📝 Contributing
We welcome contributions! See CONTRIBUTING.md for coding standards, naming conventions, and how to get started. Please open issues or pull requests for bugs, feature requests, or documentation improvements.
📰 Changelog
See CHANGELOG.md for a summary of recent changes and new features.
