doc/code/memory/chat_message.ipynb

{
 "cells": [
  {
   "cell_type": "markdown",
   "id": "0",
   "metadata": {
    "lines_to_next_cell": 0
   },
   "source": [
    "# Chat messages - optional\n",
    "\n",
    "This notebook gives an introduction to the concept of `ChatMessage` and `ChatMessageNormalizer` and how it can be helpful as you start to work with different models.\n",
    "\n",
    "\n",
    "The main format PyRIT works with is the `PromptRequestPiece` paradigm. Any time a user wants to store or retrieve a chat message, they will use the `PromptRequestPiece` object. However, `ChatMessage` is very common, so there are a lot of times this is the most useful. Any `PromptRequestPiece` object can be translated into a `ChatMessage` object.\n",
    "\n",
    "However, different models may require different formats. For example, certain models may use chatml, or may not support system messages. This is handled\n",
    "in from `ChatMessageNormalizer` and its subclasses.\n",
    "\n",
    "Below is an example that converts a list of chat messages to chatml format and back."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "1",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<|im_start|>system\n",
      "You are a helpful AI assistant<|im_end|>\n",
      "<|im_start|>user\n",
      "Hello, how are you?<|im_end|>\n",
      "<|im_start|>assistant\n",
      "I'm doing well, thanks for asking.<|im_end|>\n",
      "\n"
     ]
    }
   ],
   "source": [
    "from pyrit.chat_message_normalizer import ChatMessageNormalizerChatML\n",
    "from pyrit.models import ChatMessage\n",
    "\n",
    "messages = [\n",
    "    ChatMessage(role=\"system\", content=\"You are a helpful AI assistant\"),\n",
    "    ChatMessage(role=\"user\", content=\"Hello, how are you?\"),\n",
    "    ChatMessage(role=\"assistant\", content=\"I'm doing well, thanks for asking.\"),\n",
    "]\n",
    "\n",
    "normalizer = ChatMessageNormalizerChatML()\n",
    "chatml_messages = normalizer.normalize(messages)\n",
    "# chatml_messages is a string in chatml format\n",
    "\n",
    "print(chatml_messages)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "2",
   "metadata": {},
   "source": [
    "\n",
    "If you wish you load a chatml-format conversation, you can use the `from_chatml` method in the `ChatMessageNormalizerChatML`. This will return a list of `ChatMessage` objects that you can then use."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "3",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "[ChatMessage(role='system', content='You are a helpful AI assistant', name=None, tool_calls=None, tool_call_id=None), ChatMessage(role='user', content='Hello, how are you?', name=None, tool_calls=None, tool_call_id=None), ChatMessage(role='assistant', content=\"I'm doing well, thanks for asking.\", name=None, tool_calls=None, tool_call_id=None)]\n"
     ]
    }
   ],
   "source": [
    "chat_messages = normalizer.from_chatml(\n",
    "    \"\"\"\\\n",
    "    <|im_start|>system\n",
    "    You are a helpful AI assistant<|im_end|>\n",
    "    <|im_start|>user\n",
    "    Hello, how are you?<|im_end|>\n",
    "    <|im_start|>assistant\n",
    "    I'm doing well, thanks for asking.<|im_end|>\"\"\"\n",
    ")\n",
    "\n",
    "print(chat_messages)"
   ]
  },
  {
   "cell_type": "markdown",
   "id": "4",
   "metadata": {},
   "source": [
    "To see how to use this in action, check out the [aml endpoint](../targets/3_non_open_ai_chat_targets.ipynb) notebook. It takes a `chat_message_normalizer` parameter so that an AML model can support various chat message formats."
   ]
  },
  {
   "cell_type": "markdown",
   "id": "5",
   "metadata": {
    "lines_to_next_cell": 0
   },
   "source": [
    "Besides chatml, there are many other chat templates that a model might be trained on. If you would like to apply the template stored in a Hugging Face tokenizer,\n",
    "you can utilize `ChatMessageNormalizerTokenizerTemplate`. In the example below, we load the tokenizer for Mistral-7B-Instruct-v0.1 and apply its chat template to\n",
    "the messages. Note that this template only adds `[INST]` and `[/INST]` tokens to the user messages for instruction fine-tuning."
   ]
  },
  {
   "cell_type": "code",
   "execution_count": null,
   "id": "6",
   "metadata": {},
   "outputs": [
    {
     "name": "stdout",
     "output_type": "stream",
     "text": [
      "<s> [INST] Hello, how are you? [/INST] I'm doing well, thanks for asking.</s> [INST] What is your favorite food? [/INST]\n"
     ]
    }
   ],
   "source": [
    "import os\n",
    "\n",
    "from transformers import AutoTokenizer\n",
    "\n",
    "from pyrit.chat_message_normalizer import ChatMessageNormalizerTokenizerTemplate\n",
    "\n",
    "messages = [\n",
    "    ChatMessage(role=\"user\", content=\"Hello, how are you?\"),\n",
    "    ChatMessage(role=\"assistant\", content=\"I'm doing well, thanks for asking.\"),\n",
    "    ChatMessage(role=\"user\", content=\"What is your favorite food?\"),\n",
    "]\n",
    "\n",
    "# Load the tokenizer. If you are not logged in via CLI (huggingface-cli login), you can pass in your access token here\n",
    "# via the HUGGINGFACE_TOKEN environment variable to access the gated model.\n",
    "tokenizer = AutoTokenizer.from_pretrained(\"mistralai/Mistral-7B-Instruct-v0.1\",\n",
    "                                          token = os.environ.get(\"HUGGINGFACE_TOKEN\"))\n",
    "\n",
    "# create the normalizer and pass in the tokenizer\n",
    "tokenizer_normalizer = ChatMessageNormalizerTokenizerTemplate(tokenizer)\n",
    "\n",
    "tokenizer_template_messages = tokenizer_normalizer.normalize(messages)\n",
    "print(tokenizer_template_messages)"
   ]
  }
 ],
 "metadata": {
  "jupytext": {
   "cell_metadata_filter": "-all"
  },
  "language_info": {
   "codemirror_mode": {
    "name": "ipython",
    "version": 3
   },
   "file_extension": ".py",
   "mimetype": "text/x-python",
   "name": "python",
   "nbconvert_exporter": "python",
   "pygments_lexer": "ipython3",
   "version": "3.11.9"
  }
 },
 "nbformat": 4,
 "nbformat_minor": 5
}