LM Studio

[MadBomber]

I have been switching over to using LM Studio slowly repacing my use of ollama. I've had problems with RubyLLM::Provider using LM Studio because its responses are not 100% OpenAI compliant. I finally did this:

# lib/extensions/ruby_llm/provider_fix.rb
#
# Monkey patch to fix LM Studio compatibility with RubyLLM Provider
# LM Studio sometimes returns response.body as a String that fails JSON parsing
# This causes "String does not have #dig method" errors in parse_error

module RubyLLM
  class Provider
    # Override the parse_error method to handle String responses from LM Studio
    def parse_error(response)
      return if response.body.empty?

      body = try_parse_json(response.body)

      # Be more explicit about type checking to prevent String#dig errors
      case body
      when Hash
        # Only call dig if we're certain it's a Hash
        body.dig('error', 'message')
      when Array
        # Only call dig on array elements if they're Hashes
        body.filter_map do |part|
          part.is_a?(Hash) ? part.dig('error', 'message') : part.to_s
        end.join('. ')
      else
        # For Strings or any other type, convert to string
        body.to_s
      end
    rescue StandardError => e
      # Fallback in case anything goes wrong
      "Error parsing response: #{e.message}"
    end
  end
end

[MadBomber]

Here is the next version of the code ....

# Parses error response from provider API.
#
# Supports two error formats:
# 1. OpenAI standard: {"error": {"message": "...", "type": "...", "code": "..."}}
# 2. Simple format: {"error": "error message"}
#
# @param response [Faraday::Response] The HTTP response
# @return [String, nil] The error message or nil if parsing fails
#
# @example OpenAI format
#   response = double(body: '{"error": {"message": "Rate limit exceeded"}}')
#   parse_error(response) #=> "Rate limit exceeded"
#
# @example Simple format (LM Studio, some local providers)
#   response = double(body: '{"error": "Token limit exceeded"}')
#   parse_error(response) #=> "Token limit exceeded"
def parse_error(response)
  return if response.body.empty?

  body = try_parse_json(response.body)
  case body
  when Hash
    # Handle both formats:
    # - {"error": "message"}          (LM Studio, some providers)
    # - {"error": {"message": "..."}} (OpenAI standard)
    error_value = body['error']
    return nil unless error_value

    case error_value
    when Hash
      error_value['message']
    when String
      error_value
    else
      error_value.to_s if error_value
    end
  when Array
    body.filter_map do |part|
      next unless part.is_a?(Hash)

      error_value = part['error']
      next unless error_value

      case error_value
      when Hash then error_value['message']
      when String then error_value
      else error_value.to_s if error_value
      end
    end.join('. ')
  else
    body.to_s
  end
rescue StandardError => e
  RubyLLM.logger.debug "Error parsing response: #{e.message}"
  nil
end

[MadBomber]

@crmne I'm sitting on a PR within my fork of your library. Would you like a bug issue and a PR for this fix?

https://github.com/MadBomber/ruby_llm/pull/1/files