AWS Doc QA

AI-powered Document Q&A system. Upload PDFs, extract text via AWS Textract, and ask questions powered by Claude (Haiku 4.5) on AWS Bedrock.

Architecture

                         ┌──────────────────┐
                         │   API Gateway    │
                         │ POST /upload     │
                         │ POST /query      │
                         │ GET  /documents  │
                         └────────┬─────────┘
                                  │
              ┌───────────────────┼───────────────────┐
              ▼                   ▼                    ▼
     ┌────────────────┐  ┌───────────────┐  ┌─────────────────┐
     │ upload_handler │  │ query_handler │  │  query_handler  │
     │    Lambda      │  │   Lambda      │  │  (GET status)   │
     └───────┬────────┘  └───────┬───────┘  └────────┬────────┘
             │                   │                    │
             ▼                   ▼                    ▼
     ┌──────────────┐    ┌─────────────┐     ┌────────────────┐
     │  S3 (PDFs)   │    │ AWS Bedrock │     │   DynamoDB     │
     │  presigned   │    │  (Claude)   │     │  (status)      │
     └──────┬───────┘    └─────────────┘     └────────────────┘
            │ S3 Event
            ▼
     ┌────────────────────┐
     │ document_processor │
     │      Lambda        │
     └────────┬───────────┘
              │
              ▼
     ┌────────────────┐
     │  AWS Textract  │
     │  (async job)   │
     └────────┬───────┘
              │ SNS notify
              ▼
     ┌────────────────────┐
     │ textract_callback  │
     │      Lambda        │
     └────────┬───────────┘
              │
              ▼
     ┌────────────────┐
     │   DynamoDB     │◀──── (read by query_handler)
     │  (documents)   │
     └────────────────┘

Flow

1. Frontend sends POST /upload with filename → receives presigned S3 URL + document ID
2. Frontend uploads PDF directly to S3 via presigned URL
3. S3 event triggers document_processor Lambda → starts async Textract job
4. Textract extracts text and notifies via SNS
5. textract_callback Lambda stores extracted text in DynamoDB (status: READY)
6. Frontend polls GET /documents?document_id=... until status is READY
7. User asks a question via POST /query → query_handler retrieves text from DynamoDB, sends to Claude via Bedrock, returns the answer

API Endpoints

POST /upload

Generate a presigned S3 URL for PDF upload.

curl -X POST https://<API_ID>.execute-api.eu-central-1.amazonaws.com/prod/upload \
  -H "Content-Type: application/json" \
  -d '{"filename": "document.pdf"}'

Response:

{"upload_url": "https://s3.amazonaws.com/...", "document_id": "<uuid>_document.pdf"}

Then upload the file using the presigned URL:

curl -X PUT "<upload_url>" \
  -H "Content-Type: application/pdf" \
  --data-binary @document.pdf

GET /documents

Check document processing status.

curl "https://<API_ID>.execute-api.eu-central-1.amazonaws.com/prod/documents?document_id=<document_id>"

Response:

{"document_id": "...", "status": "PROCESSING|READY|FAILED"}

POST /query

Ask a question about a processed document.

curl -X POST https://<API_ID>.execute-api.eu-central-1.amazonaws.com/prod/query \
  -H "Content-Type: application/json" \
  -d '{"document_id": "<document_id>", "question": "What is this document about?"}'

Response:

{"document_id": "...", "question": "...", "answer": "..."}

Frontend

Single-page application built with vanilla HTML/CSS/JavaScript (no build step required).

• Two-panel layout: upload & questions on the left, answers on the right
• Drag-and-drop PDF upload with presigned URL flow
• Real-time document status polling during Textract processing
• Animated "thinking" state during Claude inference
• Response time display
• Dark/light theme support
• Mobile-responsive design

Serve frontend/index.html with any static file server or host on S3/CloudFront.

Prerequisites

• AWS CLI configured with credentials
• Python 3.12+
• Node.js (for AWS CDK CLI)
• AWS CDK CLI: npm install -g aws-cdk
• Enable Claude model access in AWS Bedrock console (eu-central-1)

Setup

# Create virtual environment
python -m venv .venv
source .venv/bin/activate  # Linux/Mac
# .venv\Scripts\activate   # Windows

# Install dependencies
pip install -r requirements-dev.txt

# Build Lambda layer
pip install -r requirements.txt -t lambdas/layer/python/

# Bootstrap CDK (first time only)
cd cdk
cdk bootstrap

# Deploy
cdk deploy

Testing

pytest tests/

Project Structure

aws-doc-qa/
├── frontend/
│   └── index.html              # Single-page web UI
├── lambdas/
│   ├── document_processor/     # S3 trigger → Textract
│   ├── textract_callback/      # Textract result → DynamoDB
│   ├── query_handler/          # API Gateway → Bedrock Claude + status polling
│   ├── upload_handler/         # Presigned S3 URL generation
│   └── layer/                  # Shared Lambda layer
├── cdk/                        # AWS CDK infrastructure
├── tests/                      # Unit tests
├── requirements.txt            # Lambda runtime dependencies
└── requirements-dev.txt        # Dev/CDK dependencies

AWS Services

Service	Purpose
S3	PDF storage (eu-central-1)
Lambda	Python 3.12, 4 functions
Textract	Async document text extraction
Bedrock	Claude Haiku 4.5 for Q&A (cross-region inference)
DynamoDB	Document metadata + extracted text
API Gateway	REST API with CORS
SNS	Textract job completion notifications