Is there a recommended way of parsing the content of messages?

[diegoquiroz]

Hey, first off, I really appreciate this library. It's by far the easiest to use for LLMs in Rails. Thank you!

I’ve noticed that messages can have three different content formats:

1. Ruby hash: "{:message=>\"Hey, how many transactions were processed this month?\", :with=>{:company_name=>\"AI company\", :user_name=>\"C3PO\"}}"
2. Ruby hash with string keys: "{\"result\"=>\"48 transactions were processed this month.\", \"query\"=>\"How many transactions were processed this month?\"}"
3. Plain string: "OK. 48 transactions were processed this month.\n"

Is there a recommended way to handle these formats when reading messages?

[diegoquiroz]

Not the most beautiful solution, but ended up with this to avoid using eval. It will save the value of :message to a new field message and all the app will work around that field instead of relying on message.content.

# Migration
add_column :messages, :message, :string

# Model
class Message < ApplicationRecord
  acts_as_message

  before_save :populate_message, if: -> { role == "user" || role == "assistant" && content.present? }

  private

  def populate_message
    parsed_content = content.gsub("=>", ":").gsub(/:(\w+)/, '"\1"')
    parsed_content = JSON.parse(parsed_content)
    self.message = parsed_content[:message] || parsed_content["message"] if parsed_content.is_a?(Hash)
  rescue SyntaxError, StandardError
      self.message = content
  end
end

[crmne]

Hey @diegoquiroz thank you for the kind words!

Just to understand better what's happening, can you make a contrived example of how you got these different messages?

[diegoquiroz]

Basically the library is sending to the LLM an object with the named attributes of ask() method, therefore the value stored in message.content could be a Ruby Hash in a string.

Example

chat = Chat.create!(
  model_id: "gemini-2.0-flash-001"
)

Just sending a string

chat.ask("Hello there!")

Results in this record

id	role	content
1	user	Hello there!

Sending string with named attribute

chat.ask(message: "Hello there!")

Results in this record

id	role	content
1	user	"{:message: "Hello there!"}"

Sending with results in the same

chat.ask(message: "Hello there!", with: { user_name: @user.name })

Results in this record

id	role	content
1	user	"{:message: "Hello there!", :with: { :user_name: "Carmine" }"

[diegoquiroz]

As I don't want to parse that hash into JSON every time I return the chat messages to my client I decided to create a column message inside Message model and store there only the content of :message.

The last example looks like this in my DB

id	role	message	content
1	user	Hello there!	"{:message: "Hello there!", :with: { :user_name: "Carmine" }"

[crmne]

Looking at your examples, I think you're running into an issue because you're using the library in a way that wasn't intended.

The ask() method is designed to take a string as its first positional argument, not a set of named parameters. The API is meant to be used like this:

# Correct usage:
chat.ask("Hello there!")

# With additional parameters:
chat.ask("What's in this image?", with: { image: "path/to/image.jpg" })

When you do:

chat.ask(message: "Hello there!")

You're actually passing a Ruby hash as the first argument, which gets converted to a string representation and stored directly. That's why you're seeing Ruby hash syntax in your content field.

The only reason this doesn't raise an error is that Ruby's flexible parameter handling quietly accepts the hash as an argument, which is only possible when chat is an ActiveRecord model btw, using a RubyLLM.chat won't allow you to do that.

This isn't a supported pattern though - it's just coincidentally not failing when used with the Rails integration.

I'd recommend sticking with the documented API pattern (message as first positional argument) going forward. No content parsing required - you'll just get plain strings as expected.

If you need to clean up existing records with serialized hashes, your workaround makes sense, though I'd simplify it and just use a regex to extract the message value.

Let me know if this clears things up!