This demo uses the example code available from OpenAI.
This is a Python script using OpenAI that uses a natural language prompt to demonstrate the OpenAI function calling capabilities.
The example at OpenAI has mock weather api calls, this version has amended the example code to use real-time OpenWeather api calls and a formatting request has been added.
This demo is available as a Docker image:
docker pull rbenson789/demo_openai_weatherdocker run -i rbenson789/demo_openai_weatherTwo api keys are required. One for OpenAI and the other for OpenWeather. This will require setting up an account for each.
Check Virtual Environment, make sure to activate.
pip install python_dotenvpip install --upgrade openaipython main.pyRAG allows real-time data and other data, that would not ordinarily be available to the LLM, to be incorporated into model generated responses.
New information, just released into the public domain will have missed the last LLM build.
There may also be valid reasons to keep information out of the public domain:
- Data regulations e.g. GDPR regulations introduced by the European Union in 2018 have reduced the amount of personal information available in the public domain.
- Competitive advantage: Research and Development
- Confidentiality: finance, defense , medical are example sectors that would keep information confidential.
- Security, govt agencies etc. will have confidential information.
Documents can be uploaded to the model using "file search". Information from these documents (PDFs, Word files, text files etc) can be used to ground responses and sources will be credited.
The LLM model can only work with data available to it. If it doesn't know about it, it will not be part of the build.
However, function calling, can make data available to the model, that would not have ordinarily been available to it.
The retrieved information, using predefined api's, is used to "augment" the model data to give more detailed, grounded responses.
For example, a car manufacturer has released a new car since the last LLM build and wants to make that information available. This can now be achieved with function calling.
The example OpenAI Code has been modified to:
- change the content for my use case, add actual real-time weather (openweather)
- change the instructions , request additional information and format the output
- add a loop to allow repeated prompt requests
- make a Docker Hub repository
rbenson789/demo_openai_weather
The repository on Docker Hub can be pulled and run on a local machine.
docker pull rbenson789/demo_openai_weatherdocker run -i rbenson789/demo_openai_weatherWeather data is supplied by openWeather api calls.
The user is invited to enter a natural language prompt, free format, with suggested guidelines.
OpenAI will determine the model arguments for building JSON for calls to a weather api. The setting up, and processing the return data will be handled by the model. The actual api call itself is a predefined api. In this case, openweather api calls are used to augment model data.
Any other weather agency could be used to fulfil the user requirements.
In the example code, 3 locations are used in the prompt, detected by the model, 3 api calls return data to the model.
Using the information in the user prompt, the model will make a response using context and inferences.
| # | Comment |
|---|---|
| Persona | Give the model an idea of the type of person that you would like a response from. e.g.You are a weather bot, you are a maths expert etc |
| Be Clear and concise | no unnecessary commentary to distract |
| Context | e.g. explain the theory of relativity |
| Scope | define boundaries, in 200 words .. |
| Specific | avoid tell me, use What ... |
| Simple | avoid jargon, technical phrases |
| Examples | en for English, nl (Dutch), fr (French) |
| Break Down | First, explain ... , then .... |
| Format | e.g. Report in table format |
| Grammar and spelling | be accurate |
| Feasible | Is the question being asked realistic and achievable ? |
The word usage is in italics. There is no strictly correct Usage in the traditional sense.
OpenAI with function calling does not operate in a strictly defined structure of semantics and arguments.
Using the prompt guidelines, enter a natural language prompt along the lines of:
What's the weather like in <location 1> <location 2> [<location>...] [% chance of rain] [language] [metric | imperial]this will also work, but be prepared for some interesting results.
<location 1> <location 2> [<location>...]When the prompt is vague, lacks context, is an unsolvable question , the model may generate a response that includes incorrect or fabricated information.
This is called hallucination.
Enter random words for the prompt, expect a random answer.
A random response is "technically correct" in that it aligns with how the model is designed to work.
However, the goal is to provide accurate and reliable information.
horseThis prompt does not meet many of the guideline criteria. A random response should therefore be expected.
This is the response for horse.
| Metric | Horse Location 🌧️ |
|-----------------|-------------------------------|
| Description | Partly cloudy with rain |
| Temperature | 15.57 °C |
| UV Index | 5.26 |
| Humidity | 55% |
| Wind Speed | 18.324 km/h |
| Wind Direction | 293° |
| Latitude | 51.9257 |
| Longitude | -1.3704 |
**Precautions for UV Index 5.26:**
- Wear sunglasses with UV protection to safeguard your eyes.
- Use sunscreen with SPF 30+ to protect exposed skin.
- Wear a hat to shield your face and neck.
- Stay in the shade during midday hours when the sun is strongest.
- Seek protective clothing like long sleeves and pants where possible.
Stay safe and enjoy your day!
The JSON generated by the model gave a latitude and longitude for horse as:
{"latitude":"53.2794","longitude":"-7.062","language":"en","unit":"metric"}this location is in the midlands of Ireland and has no obvious connection to the word horse.
The better the prompt, the better the output. A poor prompt will give a poor response.
The assistant requires 3 inputs:
- The model version, in this case gpt-4o
- Functions to run
- Instructions
This gives the model:
- A persona to take, in this case, it is a weather bot.
- directions on how to run, i.e. use functions to answer questions.
- The uv index on its own is not that meaningful. The model is told to refine that answer to add guidelines re sun exposure.
- Add an emoji
You are a weather bot.
Use the provided functions to answer questions.
Add a detailed description of the precautions to take for that particular uvi.
All output should be in requested language.
Add an emoji for the weather eg sun for sunshine.
The output should be in a table format with metrics down and locations acrossFunction calling uses defined functions that the model calls to retrieve relevant data. The functions and respective arguments are defined. The model determines the function to call by the information available to it, using input, context and inference.
If the prompt is asking for weather, it will call the get_current_weather function. If the prompt is looking for probability of rain, it will call the get_rain_probability function.
The api requests to openWeather require longitude and latitude to return JSON formatted weather results for that area.
These have not been supplied by the user nor derived within the Python code.
The model has been asked to supply these as required arguments to build JSON for api calls. By using input locations, context and inference, the model can determine these.
Creating a good prompt is important. A good prompt can significantly enhance the quality of the model's output.
Be specific and avoid unnecessary information that may distract from the main task.
This example would expect to see locations, language, percentage, units in the prompt. A vague and unspecific prompt may give vague results.
tools=[
{
"type": "function",
"function": {
"name": "get_current_weather",
"description": "Get the current weather in a given latitude and longitude",
"parameters": {
"type": "object",
"properties": {
"latitude": {
"type": "string",
"description": "The latitude of a place",
},
"longitude": {
"type": "string",
"description": "The longitude of a place",
},
"language": {
"type": "string",
"description": "The language to respond with eg en for English, nl (Dutch), fr (French), sp (Spanish), ru (Russian), uk (Ukrainian)",
},
"unit": {"type": "string",
"description": "measurements in either imperial or metric",
"enum": ["imperial", "metric"]},
},
"required": ["latitude", "longitude"],
},
},
},
{
"type": "function",
"function": {
"name": "get_rain_probability",
"description": "Get the probability of rain in a given latitude and longitude",
"parameters": {
"type": "object",
"properties": {
"latitude": {
"type": "string",
"description": "The latitude of a place",
},
"longitude": {
"type": "string",
"description": "The longitude of a place",
},
},
"required": ["latitude", "longitude"],
}
}
},
]JSON response from the get_current_weather function. The returned information from the api call will be augmented to the model. The model response will contain the following information.
result = {
"description": weather_daily_summary,
"temperature": weather_daily_temperature,
"feels_like": weather_daily_temperature_feels_like,
"humidity": weather_daily_humidity,
"wind_speed": weather_daily_wind_speed,
"wind_direction": weather_daily_wind_degree,
"latitude": latitude,
"longitude": longitude,
} What's the weather like in Sydney, Paris and Dublin, % chance of rain, in imperial units | # | Comment | |
|---|---|---|
| Persona | ✓ | instructions, assigned in Python code: assigned the role of a weather bot |
| Context | ✓ | |
| Scope | ✓ | |
| Specific | ✓ | |
| Simple | ✓ | |
| Examples | ✓ | the functions have been provided with examples en for English, nl (Dutch), fr (French) |
| Break Down | ✓ | |
| Format | ✓ | instructions, assigned in Python code: requested the output in table format etc |
| Grammar and spelling | ✓ |
Using the example prompt,the model has detected that there are 3 locations in the prompt and has extracted model arguments for api calls.
| Metric | Sydney 🌧️ | Paris 🌧️ | Dublin 🌧️ |
|---|---|---|---|
| Temperature | 59°F | 66.9°F | 57.7°F |
| Description | Partly cloudy in morning, rain in the afternoon | Partly cloudy with rain | Partly cloudy with rain |
| Wind Speed | 15.3 mph | 9.9 mph | 15 mph |
| Wind Direction | 188° | 284° | 282° |
| Humidity | 68% | 60% | 55% |
| UVI | 2.85 | 3.63 | 4.41 |
| Rain Probability | 100% | 27% | 20% |
- Sydney (UVI 2.85): Low risk. It's safe to be outside. Wear sunglasses on bright days.
- Paris (UVI 3.63): Moderate risk. Use sunscreen SPF 30+, wear sunglasses and a hat if outside for extended periods.
- Dublin (UVI 4.41): Moderate risk. Don a hat and sunglasses, and apply broad-spectrum SPF 30+ sunscreen if outside for prolonged periods.
Stay prepared and stay safe!
Prompt asking for output to be in Japanese
whats the weather like in singapore, % chance of ran , in Japanese and imperial units| メトリック | シンガポール |
|---|---|
| 天気概要 | 一時的に曇りと雨が予想されます。 |
| 気温 | 28.24°F |
| 紫外線指数 (UVI) | 10 |
| 湿度 | 78% |
| 風速 | 10.29 mph |
| 風向 | 135° |
| 降雨の確率 | 69% |
- 非常に強い紫外線に注意: なるべく外出を控え、特に午前10時から午後4時の間は注意が必要です。
- 日焼け止めを使用: 高いSPF値の日焼け止めをこまめに塗り直しましょう。
- 防護服: 長袖シャツ、長ズボン、帽子、サングラスを着用しましょう。
- 屋内活動: できるだけ屋内で活動し、直接の紫外線を避けましょう。
お出かけの際はこれらの対策を徹底してください。
In English, using Google Translate:
| Metric | Singapore |
|---|---|
| Weather summary | Partly cloudy with rain expected. |
| Temperature | 28.24°F |
| Ultraviolet index (UVI) | 10 |
| Humidity | 78% |
| Wind speed | 10.29 mph |
| Wind direction | 135° |
| Chance of rain | 69% |
- Beware of very strong UV rays: Avoid going outside as much as possible, especially between 10am and 4pm.
- Use sunscreen: Apply a high SPF sunscreen and reapply frequently.
- Protective clothing: Wear long sleeve shirts, long pants, hats and sunglasses.
- Indoor activities: Avoid direct UV rays by staying indoors as much as possible.
Please be sure to follow these precautions when going out.
FROM python:3.12.1-slim
ARG A_weather_api_key
ARG A_openai_api_key
ENV WEATHERMAP_API_KEY3=$A_weather_api_key
ENV OPENAI_API_KEY=$A_openai_api_key
# Set the working directory in the container
WORKDIR /app
# Copy the requirements.txt file into the container at /app
COPY requirements.txt .
# Install the required dependencies
RUN pip install --no-cache-dir -r requirements.txt
# Copy the rest of the application code to the working directory
COPY . .
# Command to run the script
CMD ["python", "main.py"]docker build -t weather/demo_openai_weather:1.0 .`docker build --build-arg A_weather_api_key=<your key> --build-arg A_openai_api_key=<your key> -t weather/demo_openai_weather:1.0 .The name of the image on Docker Hub will be userid/repository name
userid/repository name
e.g.
docker build --build-arg A_weather_api_key=<key> --build-arg A_openai_api_key=<key> -t rbenson789/demo_openai_weather .
docker run -i rbenson789/demo_openai_weather
Docker login
enter user id and password
docker push rbenson789/demo_openai_weather
docker pull rbenson789/demo_openai_weather
docker run -i rbenson789/demo_openai_weather
docker rm -v $(docker ps --filter status=exited -q)
The image name will be a concatenation of the user id and repository name.
userid/repository name
docker build --build-arg A_weather_api_key=<key> --build-arg A_openai_api_key=<key> -t rbenson789/demo_openai_weather .
docker login
docker push rbenson789/demo_openai_weather
docker run -i rbenson789/demo_openai_weather
docker pull rbenson789/demo_openai_weather