Table of Contents

• Retrieval Augmented Generation (RAG) Configuration - Community Edition
• Setting up RAG
  • Notes

Retrieval Augmented Generation (RAG) Configuration - Community Edition

This document applies only to the SWIRL AI Connect, Community Edition. Switch to the AI Connect, Enterprise Edition guide

SWIRL supports Real Time Retrieval Augmented Generation (RAG) out of the box, using result snippets and/or the full text of fetched result pages.

Setting up RAG

• Install SWIRL as noted in the Quick Start Guide, including the latest version of the Galaxy UI.

• Add either your OpenAI API key or your Azure OpenAI API credentials to the .env file:

OPENAI_API_KEY='your-key-here'

Check out OpenAI's YouTube video if you don't have an OpenAI API Key.

AZURE_OPENAI_KEY=<your-key>
AZURE_OPENAI_ENDPOINT=<your-azure-openai-endpoint-url>
AZURE_MODEL=<your-azure-openai-model>

SWIRL AI Connect Community Edition supports RAG only against OpenAI and Azure/OpenAI. The Enterprise Editionsupports many more.

• When installing for PRODUCTION use, change the following line in static/api/config/default from:

"webSocketConfig": {
    "url": "ws://<yourhost>:<your-port>/chatgpt-data"
  }

…to…

"webSocketConfig": {
    "url": "wss://<yourhost>:<your-port>/chatgpt-data"
  }

The default ws: prefix can be used locally but should NEVER be used in production as it is not secure!

• Add the following configuration to the page_fetch_config_json parameter of each SearchProvider you wish to have participate in RAG:

"page_fetch_config_json": {
        "cache": "false",
        "headers": {
            "User-Agent": "Swirlbot/1.0 (+http://swirl.today)"
        },
        "timeout": 10
}, 

Adjust the timeout value if necessary. Change the User-Agent string as needed, and/or authorize it to fetch pages from internal applications.

You can also override the default timeout value with a URL parameter in the Galaxy UI. For example: http://localhost:8000/galaxy/?q=gig%20economics&rag=true&rag_timeout=90

Starting with SWIRL 3.7.0, we specify rag_timeoutin seconds

• Restart SWIRL:

python swirl.py restart

• Go to the Galaxy UI (http://localhost:8000/galaxy/). The "Generate AI Response" switch should be "off" as shown:

• Run a search. Results appear quickly after you press the "Search" button (http://localhost:8000/galaxy/?q=epmc:future+of+ai+care):

• If you wish to manually select the results to RAG with, click the "Select Items" switch to make the shopping cart appear. Results that SWIRL thinks should be used in RAG will be pre-checked. Check or uncheck results, and optionally sort and/or filter them.

• Click the "Generate AI Response" switch. A spinner will appear. The RAG response will appear several seconds later depending on a variety of factors. :slightly_smiling_face:

• Verify the RAG insight you received by reviewing the citations at the end RAG response.

To cancel a RAG process, click the "Generate AI Summary" toggle off.

SWIRL's RAG processing utilizes only the first 10 results that are selected either automatically or manually using the "Select Items" option.

Notes

• The RAG process is as follows:

• The community edition of SWIRL is intended to RAG with sources you can fetch without authenticating. If you need to perform RAG with content from enterprise services like Microsoft 365, ServiceNow, Salesforce, Atlassian with OAUTH2 and SSO, please contact us for information about SWIRL Enterprise - which supports all of that, and more, out of the box.

• RAG Page fetch configurations are preloaded for the following SearchProviders:

• European PMC
• Google Web
• Google News
• LinkedIn

• SWIRL Documentation

• RAG processing is available through a single API call, e.g. ?qs=metasearch&rag=true. See the Developer Guide for more details about the ?qs= parameter.

• The default timeout value (60 seconds) and the text to display when the timeout is exceeded cab be configured in the static/api/config/default file.

"webSocketConfig": {
    "url": "ws://localhost:8000/chatgpt-data",
    "timeout": 60000,
    "timeoutText": "Timeout: No response from Generative AI."
  }