Table of Contents

  1. JSON Response Format
  2. HTML Response Format

When using Raif to interact with the LLM, you can specify the response format you want to receive. The response format will be one of text, json, or html.

You can specify the response format when chatting directly with the LLM, or when creating a Task or Conversation.

JSON Response Format

When using the json response format, Raif will:

  1. Strip any markdown code fences inserted by the LLM
  2. Parse the response as JSON.

If you’re using the OpenAI adapter, Raif will automatically enable OpenAI’s JSON mode feature.

And if you define a JSON schema, it will trigger utilization of OpenAI’s structured outputs feature, which will force the response to adhere to your schema.

HTML Response Format

When using the html response format, Raif will:

  1. Strip any markdown code fences inserted by the LLM
  2. Convert markdown links to HTML links.
  3. Strip utm_ tracking parameters from URLs.
  4. Sanitize the HTML using Rails’ sanitize method.

By default, Raif will use Rails::HTML5::SafeListSanitizer.allowed_tags and Rails::HTML5::SafeListSanitizer.allowed_attributes.

You can override these using the llm_response_allowed_tags and llm_response_allowed_tags class methods. See the Raif::Tasks::DocumentSummarization example task for usage.

Read next: Streaming Responses