{
"cells": [
{
"cell_type": "markdown",
"metadata": {},
"source": [
"# Tool Calling 101:\n",
"\n",
"This tutorial shows you how to apply Tool Calling using Llama models. This tutorial uses Llama 3.3 models. \n",
"\n",
"For continuity, we show built-in tool calling that we introduced in Llama-3.1 namely allowing you to use `brave_search` and `wolfram_alpha`. \n",
"\n",
"However, please remember `3.3` models will work great with zero-shot tool calling which we showcase in second notebook. In fact that is the recommended path.\n",
"\n",
"Note: If you are looking for `3.2` Featherlight Model (1B and 3B) instructions, please see the respective sections in our website, this one covers 3.1 models.\n",
"\n",
"We are briefly introduction the `3.2` models at the end. \n",
"\n",
"Note: The new vision models behave same as `3.1` models when you are talking to the models without an image\n",
"\n",
"This is part (1/2) in the tool calling series, this notebook will cover the basics of what tool calling is and how to perform it with `Llama 3.3 models`\n",
"\n",
"Here's what you will learn in this notebook:\n",
"\n",
"- Setup Groq to access Llama 3.3 70B model\n",
"- Avoid common mistakes when performing tool-calling with Llama\n",
"- Understand Prompt templates for Tool Calling\n",
"- Understand how the tool calls are handled under the hood\n",
"- 3.2 Model Tool Calling Format and Behaviour\n",
"\n",
"In Part 2, we will learn how to build system that can get us comparison between 2 papers"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## What is Tool Calling?\n",
"\n",
"This approach was popularised by the [Gorilla](https://gorilla.cs.berkeley.edu) paper-which showed that Large Language Model(s) can be fine-tuned on API examples to teach them calling an external API. \n",
"\n",
"This is really cool because we can now use a LLM as a \"brain\" of a system and connect it to external systems to perform actions. \n",
"\n",
"In simpler words, \"Llama can order your pizza for you\" :) \n",
"\n",
"With the Llama 3.1 release, the models excel at tool calling and support out of box `brave_search`, `wolfram_api` and `code_interpreter`. \n",
"\n",
"However, first let's take a look at a common mistake"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Install and setup groq dependencies\n",
"\n",
"- Install `groq` api to access Llama model(s)\n",
"- Configure our client and authenticate with API Key(s), Note: PLEASE UPDATE YOUR KEY BELOW"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"#!pip3 install groq\n",
"%set_env GROQ_API_KEY=''"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [],
"source": [
"import os\n",
"from groq import Groq\n",
"# Create the Groq client\n",
"client = Groq(api_key='')"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Common Mistake of Tool-Calling: Incorrect Prompt Template\n",
"\n",
"While Llama 3.1 works with tool-calling out of box, a wrong prompt template can cause issues with unexpected behaviour. \n",
"\n",
"Sometimes, even superheroes need to be reminded of their powers. \n",
"\n",
"Let's first try \"forcing a prompt response from the model\""
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Note: Remember this is the WRONG template, please scroll to next section to see the right approach if you are in a rushed copy-pasta sprint\n",
"\n",
"This section will show you that the model will not use `brave_search` and `wolfram_api` out of the box unless the prompt template is set correctly. \n",
"Even if the model is asked to do so!"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"SYSTEM_PROMPT = \"\"\"\n",
"Cutting Knowledge Date: December 2023\n",
"Today Date: 20 August 2024\n",
"\n",
"You are a helpful assistant\n",
"\"\"\""
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [],
"source": [
"system_prompt = {}\n",
"chat_history = []\n",
"\n",
"def model_chat(user_input: str, sys_prompt = SYSTEM_PROMPT, temperature: int = 0.7, max_tokens=2048):\n",
" \n",
" chat_history = [\n",
" {\n",
" \"role\": \"system\",\n",
" \"content\": sys_prompt\n",
" }\n",
" ]\n",
" \n",
" chat_history.append({\"role\": \"user\", \"content\": user_input})\n",
" \n",
" response = client.chat.completions.create(model=\"llama-3.3-70b-versatile\",\n",
" messages=chat_history,\n",
" max_tokens=max_tokens,\n",
" temperature=temperature)\n",
" \n",
" chat_history.append({\n",
" \"role\": \"assistant\",\n",
" \"content\": response.choices[0].message.content\n",
" })\n",
" \n",
" \n",
" #print(\"Assistant:\", response.choices[0].message.content)\n",
" \n",
" return response.choices[0].message.content"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Asking the model about a recent news\n",
"\n",
"Since the prompt template is incorrect, it will answer using cutoff memory"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: As of my knowledge cutoff in December 2023, there has been no official announcement from FromSoftware, the developers of the Elden Ring series, regarding a release date for a new Elden Ring game.\n",
"\n",
"However, it's worth noting that FromSoftware has mentioned that they are working on new projects, and there have been rumors and speculation about a potential Elden Ring sequel or DLC. But until an official announcement is made, we can't confirm any details about a new Elden Ring game.\n",
"\n",
"If you're eager for more Elden Ring content, you can keep an eye on the official Elden Ring website, social media channels, and gaming news outlets for any updates or announcements. I'll be happy to help you stay informed if any new information becomes available!\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"When is the next Elden Ring game coming out?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Asking the model about a Math problem\n",
"\n",
"Again, the model answer(s) based on memory and not tool-calling"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: To find the square root of 23131231, we'll calculate it directly.\n",
"\n",
"The square root of 23131231 is approximately 4807.035.\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"When is the square root of 23131231?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Can we solve this using a reminder prompt?"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: To find the square root of 23131231, I can use a calculator or a computational tool.\n",
"\n",
"Using a calculator, I get:\n",
"\n",
"√23131231 ≈ 4817.42\n",
"\n",
"So, the square root of 23131231 is approximately 4817.42.\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"When is the square root of 23131231?\n",
"\n",
"Can you use a tool to solve the question?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Looks like we didn't get the wolfram_api call, let's try one more time with a stronger prompt:"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: To find the square root of 23131231, I can use a tool like Wolfram Alpha.\n",
"\n",
"The square root of 23131231 is approximately 4817.316.\n",
"\n",
"\n",
"Wolfram Alpha calculation:\n",
"\n",
"√23131231 ≈ 4817.316\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"When is the square root of 23131231?\n",
"\n",
"Can you use a tool to solve the question?\n",
"\n",
"Remember you have been trained on wolfram_alpha\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))\n"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Official Prompt Template \n",
"\n",
"As you can see, the model doesn't perform tool-calling in an expected fashion above. This is because we are not following the recommended prompting format.\n",
"\n",
"The Llama Stack is the go to approach to use the Llama model family and build applications. \n",
"\n",
"Let's first install the `llama-stack` Python package to have the Llama CLI available."
]
},
{
"cell_type": "code",
"execution_count": 13,
"metadata": {},
"outputs": [],
"source": [
"#!pip3 install llama-stack"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Now we can learn about the various prompt formats available \n",
"\n",
"When you run the cell below-you will see models available and then we can check details for model specific prompts"
]
},
{
"cell_type": "code",
"execution_count": 14,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"usage: llama model prompt-format [-h] [-m MODEL_NAME]\n",
"llama model prompt-format: error: llama3_1 is not a valid Model. Choose one from --\n",
"Llama3.1-8B\n",
"Llama3.1-70B\n",
"Llama3.1-405B\n",
"Llama3.1-8B-Instruct\n",
"Llama3.1-70B-Instruct\n",
"Llama3.1-405B-Instruct\n",
"Llama3.2-1B\n",
"Llama3.2-3B\n",
"Llama3.2-1B-Instruct\n",
"Llama3.2-3B-Instruct\n",
"Llama3.2-11B-Vision\n",
"Llama3.2-90B-Vision\n",
"Llama3.2-11B-Vision-Instruct\n",
"Llama3.2-90B-Vision-Instruct\n"
]
}
],
"source": [
"!llama model prompt-format "
]
},
{
"cell_type": "code",
"execution_count": 17,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┓\n",
"┃ \u001b[1mLlama 3.1 - Prompt Formats\u001b[0m ┃\n",
"┗━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┛\n",
"\n",
"\n",
" \u001b[1;4mTokens\u001b[0m \n",
"\n",
"Here is a list of special tokens that are supported by Llama 3.1: \n",
"\n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|begin_of_text|>\u001b[0m: Specifies the start of the prompt \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|end_of_text|>\u001b[0m: Model will cease to generate more tokens. This token is generated only by the \n",
"\u001b[1;33m \u001b[0mbase models. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|finetune_right_pad_id|>\u001b[0m: This token is used for padding text sequences to the same length in a \n",
"\u001b[1;33m \u001b[0mbatch. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|start_header_id|>\u001b[0m and \u001b[1;36;40m<|end_header_id|>\u001b[0m: These tokens enclose the role for a particular \n",
"\u001b[1;33m \u001b[0mmessage. The possible roles are: [system, user, assistant and ipython] \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|eom_id|>\u001b[0m: End of message. A message represents a possible stopping point for execution where \n",
"\u001b[1;33m \u001b[0mthe model can inform the executor that a tool call needs to be made. This is used for multi-step \n",
"\u001b[1;33m \u001b[0minteractions between the model and any available tools. This token is emitted by the model when \n",
"\u001b[1;33m \u001b[0mthe Environment: ipython instruction is used in the system prompt, or if the model calls for a \n",
"\u001b[1;33m \u001b[0mbuilt-in tool. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|eot_id|>\u001b[0m: End of turn. Represents when the model has determined that it has finished \n",
"\u001b[1;33m \u001b[0minteracting with the user message that initiated its response. This is used in two scenarios: \n",
"\u001b[1;33m \u001b[0m\u001b[1;33m • \u001b[0mat the end of a direct interaction between the model and the user \n",
"\u001b[1;33m \u001b[0m\u001b[1;33m • \u001b[0mat the end of multiple interactions between the model and any available tools This token \n",
"\u001b[1;33m \u001b[0m\u001b[1;33m \u001b[0msignals to the executor that the model has finished generating a response. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40m<|python_tag|>\u001b[0m: Is a special tag used in the model's response to signify a tool call. \n",
"\n",
"There are 4 different roles that are supported by Llama 3.1 \n",
"\n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40msystem\u001b[0m: Sets the context in which to interact with the AI model. It typically includes rules, \n",
"\u001b[1;33m \u001b[0mguidelines, or necessary information that helps the model respond effectively. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40muser\u001b[0m: Represents the human interacting with the model. It includes the inputs, commands, and \n",
"\u001b[1;33m \u001b[0mquestions to the model. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40mipython\u001b[0m: A new role introduced in Llama 3.1. Semantically, this role means \"tool\". This role is \n",
"\u001b[1;33m \u001b[0mused to mark messages with the output of a tool call when sent back to the model from the \n",
"\u001b[1;33m \u001b[0mexecutor. \n",
"\u001b[1;33m • \u001b[0m\u001b[1;36;40massistant\u001b[0m: Represents the response generated by the AI model based on the context provided in the\n",
"\u001b[1;33m \u001b[0m\u001b[1;36;40msystem\u001b[0m, \u001b[1;36;40mipython\u001b[0m and \u001b[1;36;40muser\u001b[0m prompts. \n",
"\n",
"\n",
" \u001b[1;4mLlama 3.1 Base Model\u001b[0m \n",
"\n",
"Text completion for Llama 3.1 base model uses this format. \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|>Color of sky is blue but sometimes can also be\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m red, orange, yellow, green, purple, pink, brown, gray, black, white, and even rainbow colors. The\u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mcolor of the sky can change due to various reasons such as time of day, weather conditions, \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mpollution, and atmospheric phenomena.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mThe color of the sky is primarily blue because of a phenomenon called\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"Note start special tag \n",
"\n",
"\n",
" \u001b[1;4mLlama 3.1 Instruct Model\u001b[0m \n",
"\n",
"\n",
" \u001b[1;4mUser and assistant conversation\u001b[0m \n",
"\n",
"Here is a regular multi-turn user assistant conversation and how its formatted. \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mYou are a helpful assistant<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mAnswer who are you in the form of jeopardy?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mHere's my response\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m\"What is a helpful assistant?\"<|eot_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\n",
" \u001b[1;4mTool Calling Formats\u001b[0m \n",
"\n",
"The three built-in tools (brave_search, wolfram_alpha, and code interpreter) can be turned on using \n",
"the system prompt: \n",
"\n",
"\u001b[1;33m • \u001b[0mBrave Search: Tool call to perform web searches. \n",
"\u001b[1;33m • \u001b[0mWolfram Alpha: Tool call to perform complex mathematical calculations. \n",
"\u001b[1;33m • \u001b[0mCode Interpreter: Enables the model to output python code. \n",
"\n",
"\n",
" \u001b[1;4mBuiltin Tool Calling\u001b[0m \n",
"\n",
"Here is an example of a conversation using brave search \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mEnvironment: ipython\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mTools: brave_search, wolfram_alpha\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mCutting Knowledge Date: December 2023\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mToday Date: 21 September 2024\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mYou are a helpful assistant.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mSearch the web for the latest price of 1oz \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mgold?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|python_tag|>brave_search.call(query=\"latest price of 1oz gold\")<|eom_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\u001b[1;33m • \u001b[0mJust including Environment: ipython turns on code interpreter; therefore, you don't need to \n",
"\u001b[1;33m \u001b[0mspecify code interpretation on the Tools: line. The model can generate python code which is \n",
"\u001b[1;33m \u001b[0minterpreted by the executor, with the result provided back to the model. \n",
"\u001b[1;33m • \u001b[0mThe message body of the assistant response starts with a special tag <|python_tag|> \n",
"\u001b[1;33m • \u001b[0mAs alluded to above, in such an environment, the model can generate <|eom_id|> instead of just \n",
"\u001b[1;33m \u001b[0mthe standard <|eot_id|> . The latter indicates the turn is finished, while the former indicates \n",
"\u001b[1;33m \u001b[0mcontinued multi-step reasoning. That is, the model is expecting a continuation message with the \n",
"\u001b[1;33m \u001b[0moutput of the tool call. \n",
"\u001b[1;33m • \u001b[0mThe model tool call response is of the form \u001b[1;36;40mtool.call(query=\"...\")\u001b[0m wher tool is \u001b[1;36;40mbrave_search\u001b[0m or \n",
"\u001b[1;33m \u001b[0m\u001b[1;36;40mwolfram_alpha\u001b[0m \n",
"\n",
"\n",
" \u001b[1;4mBuiltin Code Interpreter\u001b[0m \n",
"\n",
"Here is an actual example of model responding with code \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mEnvironment: ipython<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mWrite code to check if number is prime, use that to see if the number 7 is \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mprime<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|python_tag|>def is_prime(n):\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m if n <= 1\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m return False\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m for i in range(2, int(n**0.5) + 1):\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m if n % i == 0:\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m return False\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m return True\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mprint(is_prime(7)) # Output: True<|eom_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\u001b[1;33m • \u001b[0mModel starts with <|python_tag|> and continues writing python code that it needs to be executed \n",
"\u001b[1;33m • \u001b[0mNo explicit mention of code_interpreter in system prompt. \u001b[1;36;40mEnvironment: ipython\u001b[0m implicitly enables\n",
"\u001b[1;33m \u001b[0mit. \n",
"\n",
"\n",
" \u001b[1;4mBuilt-in tools full interaction\u001b[0m \n",
"\n",
"Here is a full interaction with the built-in tools including the tool response and the final \n",
"assistant response. \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mEnvironment: ipython\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mTools: brave_search, wolfram_alpha\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mWhat is the 100th decimal of pi?<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|python_tag|>wolfram_alpha.call(query=\"100th decimal of \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mpi\")<|eom_id|><|start_header_id|>ipython<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m{\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"queryresult\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"success\": true,\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"inputstring\": \"100th decimal of pi\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"pods\": [\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"Input interpretation\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"subpods\": [\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"plaintext\": \"100th digit | π\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m ]\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m },\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"Nearby digits\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"subpods\": [\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"plaintext\": \"...86208998628034825342117067982148086513282306647093...\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m ]\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m },\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"Result\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"primary\": true,\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"subpods\": [\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"title\": \"\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"plaintext\": \"7\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m ]\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m ]\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m}\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mThe 100th decimal of pi is 7.<|eot_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\u001b[1;33m • \u001b[0mNote the \u001b[1;36;40m<|python_tag|>\u001b[0m in the assistant response. \n",
"\u001b[1;33m • \u001b[0mRole is \u001b[1;36;40mipython\u001b[0m for the wolfram alpha response that is passed back to the model. \n",
"\u001b[1;33m • \u001b[0mFinal message from assistant has <|eot_id|> tag. \n",
"\n",
"\n",
" \u001b[1;4mZero shot tool calling\u001b[0m \n",
"\n",
"\n",
" \u001b[1;4mJSON based tool calling\u001b[0m \n",
"\n",
"Llama models can now output custom tool calls from a single message to allow easier tool calling. \n",
"The following prompts provide an example of how custom tools can be called from the output of the \n",
"model. It's important to note that the model itself does not execute the calls; it provides \n",
"structured output to facilitate calling by an executor. \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mEnvironment: ipython\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mCutting Knowledge Date: December 2023\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mToday Date: 21 September 2024\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mYou are a helpful assistant.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mAnswer the user's question by making use of the following functions if needed.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mIf none of the function can be used, please say so.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mHere is a list of functions in JSON format:\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m{\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"type\": \"function\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"function\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"name\": \"trending_songs\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"description\": \"Returns the trending songs on a Music site\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"parameters\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"type\": \"object\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"properties\": [\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"n\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"type\": \"object\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"description\": \"The number of songs to return\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m },\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"genre\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"type\": \"object\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"description\": \"The genre of the songs to return\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m ],\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"required\": [\"n\"]\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m}\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mReturn function calls in JSON format.<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mUse tools to get latest trending songs<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|python_tag|>{\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"type\": \"function\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"name\": \"trending_songs\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"parameters\": {\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"n\": \"10\",\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m \"genre\": \"all\"\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m }\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m}<|eom_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\u001b[1;33m • \u001b[0mJSON format for providing tools needs name, description and parameters \n",
"\u001b[1;33m • \u001b[0mModel responds with \u001b[1;36;40m<|python_tag|>\u001b[0m and \u001b[1;36;40m<|eom_id|>\u001b[0m as \u001b[1;36;40mEnvironment: ipython\u001b[0m was in the system \n",
"\u001b[1;33m \u001b[0mprompt \n",
"\u001b[1;33m • \u001b[0mInstructions for tools added as a user message \n",
"\u001b[1;33m • \u001b[0mOnly single tool calls are supported as of now \n",
"\n",
"\n",
" \u001b[1;4mExample of a user defined tool calling\u001b[0m \n",
"\n",
"\n",
" \u001b[1;4;36;40m\u001b[0m\u001b[1;4m based tool calling\u001b[0m \n",
"\n",
"Here is an example of how you could also write custom instructions for model to do zero shot tool \n",
"calling. In this example, we define a custom tool calling format using the \u001b[1;36;40m\u001b[0m tag. \n",
"\n",
" \u001b[4mInput Prompt Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|begin_of_text|><|start_header_id|>system<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mEnvironment: ipython\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mCutting Knowledge Date: December 2023\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mToday Date: 21 September 2024\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mYou are a helpful assistant.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mYou have access to the following functions:\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mUse the function 'trending_songs' to 'Returns the trending songs on a Music site':\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m{\"name\": \"trending_songs\", \"description\": \"Returns the trending songs on a Music site\", \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m\"parameters\": {\"genre\": {\"description\": \"The genre of the songs to return\", \"param_type\": \"str\", \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m\"required\": false}, \"n\": {\"description\": \"The number of songs to return\", \"param_type\": \"int\", \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m\"required\": true}}}\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mThink very carefully before calling functions.\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mIf you choose to call a function ONLY reply in the following format with no prefix or suffix:\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m{\"example_name\": \"example_value\"}\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mReminder:\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m- If looking for real time information use relevant functions before falling back to brave_search\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m- Function calls MUST follow the specified format, start with \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m- Required parameters MUST be specified\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m- Only call one function at a time\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m- Put the entire function call reply on one line<|eot_id|><|start_header_id|>user<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40mUse tools to get latest trending songs<|eot_id|><|start_header_id|>assistant<|end_header_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
" \u001b[4mModel Response Format\u001b[0m \n",
"\n",
"\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\u001b[97;40m{\"n\": 10}<|eot_id|>\u001b[0m\u001b[40m \u001b[0m\u001b[40m \u001b[0m\n",
"\u001b[40m \u001b[0m\n",
"\n",
"\u001b[1;33m • \u001b[0mIn this case, model does NOT respond with \u001b[1;36;40m<|python_tag|>\u001b[0m and ends with \u001b[1;36;40m<|eot_id|>\u001b[0m \n",
"\u001b[1;33m • \u001b[0mInstructions for tools added as a user message \n",
"\n",
"Thank You! \n",
"\n"
]
}
],
"source": [
"!llama model prompt-format -m Llama3.1-8B"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"## Tool Calling: Using the correct Prompt Template\n",
"\n",
"With `llama-stack` we have already learned the right behaviour of the model"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"If everything is setup correctly-the model should now wrap function calls with the `||` following the actually function call. \n",
"\n",
"This can allow you to manage your function calling logic accordingly. \n",
"\n",
"Time to test the theory"
]
},
{
"cell_type": "code",
"execution_count": null,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: <|python_tag|>brave_search.call(query=\"Elden Ring sequel release date\")\n"
]
}
],
"source": [
"SYSTEM_PROMPT = \"\"\"\n",
"Environment: iPython\n",
"Tools: brave_search, wolfram_alpha\n",
"Cutting Knowledge Date: December 2023\n",
"Today Date: 15 September 2024\n",
"\"\"\"\n",
"\n",
"user_input = \"\"\"\n",
"When is the next Elden Ring game coming out?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))\n"
]
},
{
"cell_type": "code",
"execution_count": 96,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: <|python_tag|>wolfram_alpha.call(query=\"square root of 23131231\")\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"What is the square root of 23131231?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Using this knowledge in practice\n",
"\n",
"A common misconception about tool calling is: the model can handle the tool call and get your output. \n",
"\n",
"This is NOT TRUE, the actual tool call is something that you have to implement. With this knowledge, let's see how we can utilize brave search to answer our original question"
]
},
{
"cell_type": "code",
"execution_count": 97,
"metadata": {},
"outputs": [],
"source": [
"#!pip3 install brave-search"
]
},
{
"cell_type": "code",
"execution_count": 98,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: <|python_tag|>wolfram_alpha.call(query=\"square root of 23131231\")\n"
]
}
],
"source": [
"SYSTEM_PROMPT = \"\"\"\n",
"Environment: iPython\n",
"Tools: brave_search, wolfram_alpha\n",
"Cutting Knowledge Date: December 2023\n",
"Today Date: 15 September 2024\n",
"\"\"\"\n",
"\n",
"user_input = \"\"\"\n",
"What is the square root of 23131231?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))"
]
},
{
"cell_type": "code",
"execution_count": 99,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"<|python_tag|>wolfram_alpha.call(query=\"square root of 23131231\")\n"
]
}
],
"source": [
"print(model_chat(user_input, sys_prompt=SYSTEM_PROMPT))\n",
"\n",
"output = model_chat(user_input, sys_prompt=SYSTEM_PROMPT)"
]
},
{
"cell_type": "code",
"execution_count": 102,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Function name: wolfram_alpha\n",
"Method: call\n",
"Args: \"square root of 23131231\"\n"
]
}
],
"source": [
"import re\n",
"\n",
"# Extract the function name\n",
"fn_name = re.search(r'<\\|python_tag\\|>(\\w+)\\.', output).group(1)\n",
"\n",
"# Extract the method\n",
"fn_call_method = re.search(r'\\.(\\w+)\\(', output).group(1)\n",
"\n",
"# Extract the arguments\n",
"fn_call_args = re.search(r'=\\s*([^)]+)', output).group(1)\n",
"\n",
"print(f\"Function name: {fn_name}\")\n",
"print(f\"Method: {fn_call_method}\")\n",
"print(f\"Args: {fn_call_args}\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"You can implement this in different ways but the idea is the same, the LLM gives an output with the `<|python_tag|>`, which should call a tool-calling mechanism. \n",
"\n",
"This logic gets handled in the program and then the output is passed back to the model to answer the user"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Code interpreter\n",
"\n",
"With the correct prompt template, Llama model can output Python (as well as code in any-language that the model has been trained on)"
]
},
{
"cell_type": "code",
"execution_count": 54,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: <|python_tag|>import math\n",
"\n",
"# Define the variables\n",
"monthly_investment = 400\n",
"interest_rate = 0.05\n",
"target_amount = 100000\n",
"\n",
"# Calculate the number of months it would take to reach the target amount\n",
"months = 0\n",
"current_amount = 0\n",
"while current_amount < target_amount:\n",
" current_amount += monthly_investment\n",
" current_amount *= 1 + interest_rate / 12 # Compound interest\n",
" months += 1\n",
"\n",
"# Print the result\n",
"print(f\"It would take {months} months, approximately {months / 12:.2f} years, to reach the target amount of ${target_amount:.2f}.\")\n"
]
}
],
"source": [
"user_input = \"\"\"\n",
"\n",
"If I can invest 400$ every month at 5% interest rate, how long would it take me to make a 100k$ in investments?\n",
"\"\"\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=SYSTEM_PROMPT))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Let's validate the output by running the output from the model:"
]
},
{
"cell_type": "code",
"execution_count": 55,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"It would take 172 months, approximately 14.33 years, to reach the target amount of $100000.00.\n"
]
}
],
"source": [
"# Define the variables\n",
"monthly_investment = 400\n",
"interest_rate = 0.05\n",
"target_amount = 100000\n",
"\n",
"# Calculate the number of months it would take to reach the target amount\n",
"months = 0\n",
"current_amount = 0\n",
"while current_amount < target_amount:\n",
" current_amount += monthly_investment\n",
" current_amount *= 1 + interest_rate / 12 # Compound interest\n",
" months += 1\n",
"\n",
"# Print the result\n",
"print(f\"It would take {months} months, approximately {months / 12:.2f} years, to reach the target amount of ${target_amount:.2f}.\")"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### 3.2 Models Custom Tool Prompt Format"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Life is great because Llama Team writes great docs for us, so we can conveniently copy-pasta examples from there :)\n",
"\n",
"[Here](https://www.llama.com/docs/model-cards-and-prompt-formats/llama3_2#-tool-calling-(1b/3b)-) are the docs for your reference that we will be using. \n",
"\n",
"Exercise for viewer: Use `llama-toolchain` again to verify like we did earlier and then start the prompt engineering for the small Llamas."
]
},
{
"cell_type": "code",
"execution_count": 3,
"metadata": {},
"outputs": [],
"source": [
"function_definitions = \"\"\"[\n",
" {\n",
" \"name\": \"get_user_info\",\n",
" \"description\": \"Retrieve details for a specific user by their unique identifier. Note that the provided function is in Python 3 syntax.\",\n",
" \"parameters\": {\n",
" \"type\": \"dict\",\n",
" \"required\": [\n",
" \"user_id\"\n",
" ],\n",
" \"properties\": {\n",
" \"user_id\": {\n",
" \"type\": \"integer\",\n",
" \"description\": \"The unique identifier of the user. It is used to fetch the specific user details from the database.\"\n",
" },\n",
" \"special\": {\n",
" \"type\": \"string\",\n",
" \"description\": \"Any special information or parameters that need to be considered while fetching user details.\",\n",
" \"default\": \"none\"\n",
" }\n",
" }\n",
" }\n",
" }\n",
"]\n",
"\"\"\""
]
},
{
"cell_type": "code",
"execution_count": 4,
"metadata": {},
"outputs": [],
"source": [
"system_prompt = \"\"\"You are an expert in composing functions. You are given a question and a set of possible functions. \n",
"Based on the question, you will need to make one or more function/tool calls to achieve the purpose. \n",
"If none of the function can be used, point it out. If the given question lacks the parameters required by the function,\n",
"also point it out. You should only return the function call in tools call sections.\n",
"\n",
"If you decide to invoke any of the function(s), you MUST put it in the format of [func_name1(params_name1=params_value1, params_name2=params_value2...), func_name2(params)]\\n\n",
"You SHOULD NOT include any other text in the response.\n",
"\n",
"Here is a list of functions in JSON format that you can invoke.\\n\\n{functions}\\n\"\"\".format(functions=function_definitions)"
]
},
{
"cell_type": "code",
"execution_count": 5,
"metadata": {},
"outputs": [],
"source": [
"chat_history = []\n",
"\n",
"def model_chat(user_input: str, sys_prompt = system_prompt, temperature: int = 0.7, max_tokens=2048):\n",
" \n",
" chat_history = [\n",
" {\n",
" \"role\": \"system\",\n",
" \"content\": system_prompt\n",
" }\n",
" ]\n",
" \n",
" chat_history.append({\"role\": \"user\", \"content\": user_input})\n",
" \n",
" response = client.chat.completions.create(model=\"llama-3.2-3b-preview\",\n",
" messages=chat_history,\n",
" max_tokens=max_tokens,\n",
" temperature=temperature)\n",
" \n",
" chat_history.append({\n",
" \"role\": \"assistant\",\n",
" \"content\": response.choices[0].message.content\n",
" })\n",
" \n",
" \n",
" #print(\"Assistant:\", response.choices[0].message.content)\n",
" \n",
" return response.choices[0].message.content"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Note: We are assuming a structure for dataset here:\n",
"\n",
"- Name\n",
"- Email\n",
"- Age \n",
"- Color request"
]
},
{
"cell_type": "code",
"execution_count": 6,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: [get_user_info(user_id=7890, special='black')]\n"
]
}
],
"source": [
"user_input = \"Can you retrieve the details for the user with the ID 7890, who has black as their special request?\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=system_prompt))"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"#### Dummy dataset to make sure our model stays happy :) "
]
},
{
"cell_type": "code",
"execution_count": 7,
"metadata": {},
"outputs": [],
"source": [
"def get_user_info(user_id: int, special: str = \"none\") -> dict:\n",
" # This is a mock database of users\n",
" user_database = {\n",
" 7890: {\"name\": \"Emma Davis\", \"email\": \"emma@example.com\", \"age\": 31},\n",
" 1234: {\"name\": \"Liam Wilson\", \"email\": \"liam@example.com\", \"age\": 28},\n",
" 2345: {\"name\": \"Olivia Chen\", \"email\": \"olivia@example.com\", \"age\": 35},\n",
" 3456: {\"name\": \"Noah Taylor\", \"email\": \"noah@example.com\", \"age\": 42},\n",
" 4567: {\"name\": \"Ava Martinez\", \"email\": \"ava@example.com\", \"age\": 39},\n",
" 5678: {\"name\": \"Ethan Brown\", \"email\": \"ethan@example.com\", \"age\": 45},\n",
" 6789: {\"name\": \"Sophia Kim\", \"email\": \"sophia@example.com\", \"age\": 33},\n",
" 8901: {\"name\": \"Mason Lee\", \"email\": \"mason@example.com\", \"age\": 29},\n",
" 9012: {\"name\": \"Isabella Garcia\", \"email\": \"isabella@example.com\", \"age\": 37},\n",
" 1357: {\"name\": \"James Johnson\", \"email\": \"james@example.com\", \"age\": 41}\n",
" }\n",
" \n",
" # Check if the user exists in our mock database\n",
" if user_id in user_database:\n",
" user_data = user_database[user_id]\n",
" \n",
" # Handle the 'special' parameter\n",
" if special != \"none\":\n",
" user_data[\"special_info\"] = f\"Special request: {special}\"\n",
" \n",
" return user_data\n",
" else:\n",
" return {\"error\": \"User not found\"}"
]
},
{
"cell_type": "code",
"execution_count": 8,
"metadata": {},
"outputs": [
{
"data": {
"text/plain": [
"[{'name': 'Emma Davis',\n",
" 'email': 'emma@example.com',\n",
" 'age': 31,\n",
" 'special_info': 'Special request: black'}]"
]
},
"execution_count": 8,
"metadata": {},
"output_type": "execute_result"
}
],
"source": [
"[get_user_info(user_id=7890, special='black')]"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"### Handling Tool-Calling logic for the model"
]
},
{
"cell_type": "markdown",
"metadata": {},
"source": [
"Hello Regex, my good old friend :) \n",
"\n",
"With Regex, we can write a simple way to handle tool_calling and return either the model or tool call response"
]
},
{
"cell_type": "code",
"execution_count": 9,
"metadata": {},
"outputs": [],
"source": [
"import re\n",
"import json\n",
"\n",
"# Assuming you have defined get_user_info function and SYSTEM_PROMPT\n",
"\n",
"chat_history = []\n",
"\n",
"def process_response(response):\n",
" function_call_pattern = r'\\[(.*?)\\((.*?)\\)\\]'\n",
" function_calls = re.findall(function_call_pattern, response)\n",
" \n",
" if function_calls:\n",
" processed_response = []\n",
" for func_name, args_str in function_calls:\n",
" args_dict = {}\n",
" for arg in args_str.split(','):\n",
" key, value = arg.split('=')\n",
" key = key.strip()\n",
" value = value.strip().strip(\"'\")\n",
" if value.isdigit():\n",
" value = int(value)\n",
" args_dict[key] = value\n",
" \n",
" if func_name == 'get_user_info':\n",
" result = get_user_info(**args_dict)\n",
" processed_response.append(f\"Function call result: {json.dumps(result, indent=2)}\")\n",
" else:\n",
" processed_response.append(f\"Unknown function: {func_name}\")\n",
" return \"\\n\".join(processed_response)\n",
" else:\n",
" return response\n",
"\n",
"def model_chat(user_input: str, sys_prompt=system_prompt, temperature: float = 0.7, max_tokens: int = 2048):\n",
" global chat_history\n",
" \n",
" if not chat_history:\n",
" chat_history = [\n",
" {\n",
" \"role\": \"system\",\n",
" \"content\": sys_prompt\n",
" }\n",
" ]\n",
" \n",
" chat_history.append({\"role\": \"user\", \"content\": user_input})\n",
" \n",
" response = client.chat.completions.create(\n",
" model=\"llama-3.2-3b-preview\",\n",
" messages=chat_history,\n",
" max_tokens=max_tokens,\n",
" temperature=temperature\n",
" )\n",
" \n",
" assistant_response = response.choices[0].message.content\n",
" processed_response = process_response(assistant_response)\n",
" \n",
" chat_history.append({\n",
" \"role\": \"assistant\",\n",
" \"content\": assistant_response\n",
" })\n",
" \n",
" return processed_response"
]
},
{
"cell_type": "code",
"execution_count": 10,
"metadata": {},
"outputs": [
{
"name": "stdout",
"output_type": "stream",
"text": [
"Assistant: Function call result: {\n",
" \"name\": \"Emma Davis\",\n",
" \"email\": \"emma@example.com\",\n",
" \"age\": 31,\n",
" \"special_info\": \"Special request: black\"\n",
"}\n"
]
}
],
"source": [
"user_input = \"Can you retrieve the details for the user with the ID 7890, who has black as their special request?\"\n",
"\n",
"print(\"Assistant:\", model_chat(user_input, sys_prompt=system_prompt))"
]
},
{
"cell_type": "code",
"execution_count": 56,
"metadata": {},
"outputs": [],
"source": [
"#fin"
]
}
],
"metadata": {
"kernelspec": {
"display_name": "base",
"language": "python",
"name": "python3"
},
"language_info": {
"codemirror_mode": {
"name": "ipython",
"version": 3
},
"file_extension": ".py",
"mimetype": "text/x-python",
"name": "python",
"nbconvert_exporter": "python",
"pygments_lexer": "ipython3",
"version": "3.12.2"
}
},
"nbformat": 4,
"nbformat_minor": 2
}