Skip to main content

​
Overview

AutoRAG is Cloudflare’s fully managed RAG service - the easiest way to do retrieval-augmented generation. Just point it to an R2 bucket and it handles everything automatically:
  • ✅ Automatic document ingestion from R2 buckets
  • ✅ Automatic chunking with smart splitting
  • ✅ Automatic embedding via Workers AI
  • ✅ Automatic indexing in Vectorize
  • ✅ Continuous monitoring and updates
  • ✅ Supports PDFs, images, text, HTML, CSV, and more
This is the recommended way to do RAG on Cloudflare - zero manual work required!

​
Quick Example

name: autorag-qa
description: Q&A using AutoRAG

flow:
  - member: search-docs
    type: AutoRAG
    config:
      instance: my-knowledge-base
      mode: answer  # Returns AI-generated answer
    input:
      query: ${input.question}

output:
  answer: ${search-docs.output.answer}
  sources: ${search-docs.output.sources}

​
Setup

​
1. Configure in wrangler.toml

[[autorag]]
binding = "MY_KNOWLEDGE_BASE"
bucket = "my-docs-bucket"  # Your R2 bucket name

# Optional configuration
chunk_size = 512
chunk_overlap = 50
embedding_model = "@cf/baai/bge-base-en-v1.5"

​
2. Upload Documents to R2

# Upload your documents
wrangler r2 object put my-docs-bucket/doc1.pdf --file=./doc1.pdf
wrangler r2 object put my-docs-bucket/doc2.md --file=./doc2.md
That’s it! CloudFlare automatically:
  • Detects new files in R2
  • Extracts text content
  • Chunks documents
  • Generates embeddings
  • Indexes in Vectorize

​
Configuration

config:
  # AutoRAG instance name from wrangler.toml
  instance: string  # Required

  # Return mode: 'answer' or 'results'
  mode: 'answer' | 'results'  # Default: 'answer'

  # Number of results to retrieve
  topK: number  # Default: 5

  # Enable query rewriting for better retrieval
  rewriteQuery: boolean  # Default: false

​
Modes

Returns an AI-generated answer grounded in retrieved documents: 
- member: ask-question
  type: AutoRAG
  config:
    instance: kb
    mode: answer  # AI generates answer
  input:
    query: "What are the key features?"
Output: 
output:
  answer: "The key features include..."
  sources:
    - content: "Feature 1 is..."
      score: 0.95
      metadata: {filename: "doc1.pdf"}
    - content: "Feature 2 is..."
      score: 0.89
      metadata: {filename: "doc2.md"}

​
Results Mode

Returns raw search results without generation: 
- member: search
  type: AutoRAG
  config:
    instance: kb
    mode: results  # Just retrieve, don't generate
  input:
    query: "features"
    topK: 10
Output: 
output:
  results:
    - content: "Feature 1 description..."
      score: 0.95
      metadata: {filename: "doc1.pdf", page: 3}
    - content: "Feature 2 description..."
      score: 0.89
      metadata: {filename: "doc2.md"}

​
Supported File Types

AutoRAG automatically handles:
  • Text: .txt, .md, .csv
  • Documents: .pdf, .docx
  • Code: .js, .ts, .py, .go, etc.
  • Web: .html, .xml
  • Images: .jpg, .png (with OCR)

​
Complete Example

name: document-qa-system
description: Full Q&A system with AutoRAG

flow:
  # Search knowledge base
  - member: search
    type: AutoRAG
    config:
      instance: company-docs
      mode: answer
      topK: 5
    input:
      query: ${input.question}

  # If no good answer, escalate
  - member: check-confidence
    type: Function
    input:
      sources: ${search.output.sources}
      threshold: 0.7

  - member: escalate
    condition: ${check-confidence.output.lowConfidence}
    type: API
    config:
      url: ${env.SLACK_WEBHOOK}
      method: POST
    input:
      body:
        text: "Question needs human review: ${input.question}"

output:
  answer: ${search.output.answer}
  confidence: ${check-confidence.output.confidence}
  sources: ${search.output.sources}
  escalated: ${escalate.success || false}

​
Advantages

vs. Manual Vectorize-RAG:
  • ✅ No manual chunking code
  • ✅ No embedding generation code
  • ✅ No indexing logic
  • ✅ Automatic updates when R2 files change
  • ✅ Built-in monitoring
vs. External RAG Services:
  • ✅ No data egress - stays in Cloudflare
  • ✅ Lower latency - edge-native
  • ✅ No extra costs - bundled pricing
  • ✅ Integrated with Workers

​
Best Practices

  1. Organize R2 bucket - Use folders for categories
  2. Descriptive filenames - Used in metadata
  3. Monitor bucket size - AutoRAG has limits
  4. Use answer mode - Better UX than raw results
  5. Set appropriate topK - Balance speed vs. completeness
  6. Test queries - Verify retrieval quality

​
Limitations

  • Bucket size: Check AutoRAG pricing for limits
  • File size: Individual files have size limits
  • Update latency: New files indexed within minutes
  • Query rate: Standard Workers rate limits apply

Vectorize-RAG

Manual Vectorize integration

Think Members

AI generation

AutoRAG Docs

Official AutoRAG documentation