Extract Data with On-demand and Batch Pipelines Dynamically

2026-06-11 · Source: Artificial Intelligence · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cloud Computing & IT Infrastructure · Depth: Intermediate, long

Summary

This post details an intelligent document processing solution leveraging Amazon Bedrock for extracting data from large volumes of paper or electronic documents. The architecture features both on-demand and batch inference pipelines, offering flexibility in processing time and cost. The on-demand pipeline processes single documents rapidly for time-sensitive needs, while the batch pipeline is optimized for cost, achieving a 50% reduction compared to on-demand. A key capability is dynamically specifying the large language model and prompts at the document level, enabling accurate data extraction from diverse document types like scanned land lease PDFs. The solution utilizes AWS SQS, Lambda, S3, DynamoDB, and EventBridge, processing up to 1,000 documents within 15 minutes using Python's multiprocessing.

Key takeaway

For MLOps Engineers managing document processing workflows, this architecture offers a robust approach to balance speed and cost. You should implement a dual-pipeline strategy, using on-demand inference for critical, time-sensitive documents and batch processing for large backlogs to achieve significant cost savings, potentially 50%. Dynamically managing prompts and models via Amazon Bedrock Prompt Management will enhance extraction accuracy across diverse document types, improving overall system adaptability and efficiency.

Key insights

Amazon Bedrock enables flexible, cost-optimized intelligent document processing via dynamic on-demand and batch inference pipelines.

Principles

• Balance processing speed with cost efficiency.
• Tailor prompts dynamically for varied document formats.
• Utilize FIFO queues for ordered, reliable processing.

Method

The solution uses SQS to trigger Lambda functions that retrieve documents, convert PDFs to images, fetch tailored prompts from Bedrock Prompt Management, invoke LLMs (Converse API), and store results in DynamoDB, with separate flows for on-demand and batch.

In practice

• Process scanned PDFs with multimodal models like Claude 4 Sonnet.
• Split large documents into 20-page chunks for LLM limits.
• Use JSONL artifacts for Amazon Bedrock batch inference jobs.

Topics

• Intelligent Document Processing
• Amazon Bedrock
• Generative AI
• LLM Inference
• AWS Lambda
• Prompt Engineering

Code references

• build-on-aws/amazon-bedrock-quick-start

Best for: AI Engineer, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• Rocket Close transforms mortgage document processing with Amazon Bedrock and Amazon Textract — Generative AI, combined with OCR, can automate complex document processing with high accuracy and speed.
• Cost-efficient custom text-to-SQL using Amazon Nova Micro and Amazon Bedrock on-demand inference — Fine-tuning Amazon Nova Micro with LoRA on Amazon Bedrock provides cost-efficient, production-ready text-to-SQL for custom dialects.
• RAG in Production: Designing Retrieval Pipelines That Stay Accurate as Your Data Changes — Production RAG systems require dynamic, continuous retrieval pipelines to maintain accuracy with evolving data.
• From Manual Document Processing to AI-Orchestrated Intelligence — An AI-orchestrated pipeline automates document processing using dual-model extraction, human oversight, and zero-code extensibility.
• Serving Multiple Users at Once: How Continuous Batching Keeps LLM Inference Efficient — Continuous batching optimizes LLM inference by dynamically scheduling requests and packing tokens to eliminate padding and idle GPU time.
• 3 SpaCy Tricks for Efficient Text Processing & Entity Recognition — Optimizing spaCy pipelines requires tailoring component usage, utilizing batch processing, and integrating rule-based entity recognition.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by Artificial Intelligence.