Serverless document and media processing with AI chat. Scale-to-zero architecture — no vector database fees, no idle costs. Upload documents, images, video, and audio — extract text with OCR or transcription — query using Amazon Bedrock or your AI assistant via MCP.
QUESTIONS?
Features
- ☁️ Fully serverless architecture (Lambda, Step Functions, S3, DynamoDB)
- 🧠 NEW Amazon Nova multimodal embeddings for text and image vectorization
- 📄 Document processing & vectorization (PDF, images, Office docs, HTML, CSV, JSON, XML, EML, EPUB) → stored in managed knowledge base
- 🎬 NEW Video/audio processing - transcribe speech with AWS Transcribe, searchable by timestamp
- 💬 AI chat with retrieval-augmented context and source attribution
- 📎 Collapsible source citations with optional document downloads
- ⏱️ NEW Media sources with timestamp links - click to play at exact position
- 🔍 Metadata filtering - auto-discover document metadata and filter search results
- 🎯 Relevancy boost for filtered results - prioritize matches from metadata filters
- 🔄 Knowledge Base reindex - regenerate metadata for existing documents with updated settings
- 🗑️ Document management - reprocess, reindex, or delete documents from the dashboard
- 🌐 Web component for any framework (React, Vue, Angular, Svelte)
- 🚀 One-click deploy
- 💰 $7-10/month (1000 docs, Textract + Haiku)
Live Demo
Base Pipeline: The core document processing tool - upload, OCR, and query documents.
Project Showcase: See RAGStack powering a real application.
Quick Start
Option 1: One-Click Deploy (AWS Marketplace)
REPO IS IN ACTIVE DEVELOPMENT AND WILL CHANGE OFTEN
Deploy directly from the AWS Console - no local setup required:
- Subscribe to RAGStack on AWS Marketplace (free, not required - If subscribed Lambda roles auto-accept Bedrock model agreements on first invocation)
- Click here to deploy
- Enter a stack name (lowercase only, e.g., "my-docs") and your admin email
- Click Create Stack (deployment takes ~10 minutes)
After deployment:
- Check your email for the temporary password (from Cognito)
- Go to CloudFormation → your stack → Outputs tab to find the Dashboard URL (
UIUrl)
Option 2: Deploy from Source
For customization or development:
Prerequisites:
- AWS Account with admin access
- Python 3.13+, Node.js 24+
- uv (Python package manager)
- AWS CLI, SAM CLI (configured)
- Docker (for Lambda layer builds)
git clone https://github.com/HatmanStack/RAGStack-Lambda.git
cd RAGStack-Lambda
uv sync
python publish.py \
--stack-name my-docs \
--admin-email admin@example.com
Option 3: Nested Stack Deployment
Deploy RAGStack as part of a larger CloudFormation stack. See Nested Stack Deployment Guide for details.
Quick example:
Resources:
RAGStack:
Type: AWS::CloudFormation::Stack
Properties:
TemplateURL: https://ragstack-quicklaunch-public.s3.us-east-1.amazonaws.com/ragstack-template.yaml
Parameters:
StackPrefix: 'my-app-ragstack'
AdminEmail: admin@example.com
Web Component Integration
See RAGSTACK_CHAT.md for web component integration guide.
API Access
Server-side integrations use API key authentication. Get your key from Dashboard → Settings.
curl -X POST 'YOUR_GRAPHQL_ENDPOINT' \
-H 'x-api-key: YOUR_API_KEY' \
-H 'Content-Type: application/json' \
-d '{"query": "query { searchKnowledgeBase(query: \"...\") { results { content } } }"}'
Web component uses IAM auth (no API key needed - handled automatically).
Each UI tab shows server-side API examples in an expandable section.
MCP Server (AI Assistant Integration)
Use your knowledge base directly in Claude Desktop, Cursor, VS Code, Amazon Q CLI, and other MCP-compatible tools.
Add to your AI assistant's MCP config:
{
"ragstack-kb": {
"command": "uvx",
"args": ["ragstack-mcp"],
"env": {
"RAGSTACK_GRAPHQL_ENDPOINT": "YOUR_ENDPOINT",
"RAGSTACK_API_KEY": "YOUR_API_KEY"
}
}
}
Then ask naturally: "Search my knowledge base for authentication docs"
See MCP Server docs for full setup instructions.
Architecture
Upload → OCR → Embeddings → Bedrock KB
↓
Web UI (Dashboard + Chat) ←→ GraphQL API
↓
Web Component ←→ AI Chat with Sources
Usage
Documents
Upload documents in various formats. Auto-detection routes to optimal processor:
| Type | Formats | Processing |
|---|
| Text | HTML, TXT, CSV, JSON, XML, EML, EPUB, DOCX, XLSX | Direct extraction with smart analysis |
| OCR | PDF, JPG, PNG, TIFF, GIF, BMP, WebP, AVIF | Textract or Bedrock vision OCR (WebP/AVIF require Bedrock) |
| Media | MP4, WebM, MP3, WAV, M4A, OGG, FLAC | AWS Transcribe → 30s segments → searchable with timestamps |
| Passthrough | Markdown (.md) | Direct copy |
Processing time: UPLOADED → PROCESSING → INDEXED (typically 1-5 min for text, 2-15 min for OCR, 5-20 min for media)
Images
Upload JPG, PNG, GIF, WebP with captions. Both visual content and caption text are searchable.
Web Scraping
Scrape websites into the knowledge base. See Web Scraping.
Video & Audio
Upload MP4, WebM, MP3, WAV, M4A, OGG, or FLAC files. Speech is transcribed using AWS Transcribe and segmented into 30-second chunks for search. Sources include timestamps (e.g., "1:30-2:00") with clickable links that play at the exact position.
Features:
- Speaker diarization (identify who said what)
- Configurable language (30+ languages supported)
- Timestamp-linked sources in chat responses
See Configuration for language and speaker settings.
Chat
Ask questions about your content. Sources show where answers came from.
Documentation
Development
Deployment Options
Direct Deployment
python publish.py --stack-name myapp --admin-email admin@example.com
python publish.py --stack-name myapp --admin-email admin@example.com --skip-ui
python publish.py --stack-name myapp --admin-email admin@example.com --skip-ui-all
python publish.py --stack-name myapp --admin-email admin@example.com --demo-mode
Publish to AWS Marketplace (Maintainers)
To update the one-click deploy template:
python publish.py --publish-marketplace
This packages the application and uploads to S3 for one-click deployment.
Note: Currently requires us-east-1 (Nova Multimodal Embeddings). When available in other regions, use --region <region>.
Acknowledgments
This project was inspired by: