A command line tool for running GPT commands. This tool supports prompt-batching, prompt-chaining and ChatGPT Plugins.
Use this tool to
- Create batch processing of GPT requests. Run a set of data against a prompt and record the responses.
- Create experiments to see how different config parameters and prompts affect performance and output.
- Create images with DALL-E
- Do rapid prototyping of ChatGPT Plugins through a mock server
- Discover ChatGPT Plugins with a ChatGPT Plugin
You will need to create an OpenAI API Key. If you have an account, you can create a key here
https://platform.openai.com/account/api-keys
Do not share your API key with others, or expose it in the browser or other client-side code. You will incur any charges if someone uses your key. Don't check your key into any public repository.
Create a file that contains your API Key (the one below is not real). In our example. we name the file api_key and add the key.
sk-gKtTxOumv4orO6cfWlh0ZK
Make sure your have dart installed. Follow the instructions, in the link below.
After installation, you can install the gpt program with the following command
dart pub global activate gpt
The following are the use cases supported
- Generate ChatGPT Plugin
- Generate Image Project
- Generate Batch Project
- Generate Prompt Project
- Generate Chain Project
- Add Report Plugin
- ZapVine Plugin Catalog
Run the generate project command
air genp
You will first need to select the archetype
? Project Archetype ›
❯ Prompt
Chain
Batch
Image
ChatGPT Plugin
Enter the projectName and projectVersion
✔ Project Archetype · Prompt
✔ Project Name: · myproject
✔ Project Version: · 1.0
Depending on the project, you may need to enter your API Key. You can skip, use an existing key file or create a new key file
? Import Key ›
❯ Skip
Use Existing OpenAI API Key File
Create New OpenAI API Key File
The following option allows us to enter the api key directly. It will save the key to a file. If you have trouble copying and pasting the key, just enter a few characters and then edit the file afterwards.
✔ Import Key · Create New OpenAI API Key File
? API Key: › sk-gKtTxOumv4orO6cfWlh0ZK
The ChatGPT Plugin project allows you to do rapid prototyping of a ChatGPT Plugin. Specifically it allows you to mock responses to ChatGPT. The project is based upon the quickstart project at: https://github.com/openai/plugins-quickstart
Your project file will look like
---
projectName: plugin-quickstart
projectVersion: '1.0'
projectType: plugin
defaultConfig:
properties:
port: 5003
nameForHuman: TODO Plugin (no auth)
nameForModel: todo
descriptionForHuman: Plugin for managing a TODO list, you can add, remove and view your TODOs.
mockedRequests:
- path: "/todos/global"
method: get
mockedResponse: todos-global.json
- path: "/todos/user"
method: get
mockedResponse: todos-global.json
pluginServers:
- serverId: todo-mike
flavor: mike
# mocked requests
mockedRequests:
- path: "/todos/mike"
method: get
mockedResponse: todos.json # returns the content of this file
# Adds a different user for testing
- serverId: todo-kaleb
flavor: kaleb
mockedRequests:
- path: "/todos/kaleb"
method: get
mockedResponse: todos.json
properties:
showHttpHeaders: true # Show http headers in logsAny configuration in the defaultConfig node will be applied to each pluginServer unless that plugin specifically overrides the property.
A sample mocked response (todos.json) is given below. This will be returned on a call to /todos/mike
{
"todos": [
"Clean out a septic tank",
"Repair a broken sewer pipe",
"Collect roadkill for disposal",
"Assist in bee hive relocation",
"Service a grease trap at a restaurant"
]
}To start a mocked instance of the plugin server
air plugin
or to start a specific server add the serverId option
air plugin --serverId todo-mike
For more information about creating and using a ChatGPT-Plugin
If you chose to create an image project, you will be asked for a description. Don't worry you can change it later after generating the project.
? Image Description: › A goldfish with big eyes
Your project file will look like
projectName: image-2
projectVersion: 2.0
apiKeyFile: api_key
blocks:
- blockId: image-block-1
pluginName: ImageGptPlugin
executions:
#First Image
- id: img-1
sizes:
- 256
- 256
- 1024
prompt: image.prompt
properties:
imageDescription: A goldfish with big eyesYou prompt file will be
Generate a picture of a ${imageDescription}
For more information about images
The following asks how many times to execute the batch data. If you choose 5 times, it will run all the batch data calls 5 times each.
? Number of Times to Run The Block: › 5
You will see a project file like the following
---
projectName: mybatch
projectVersion: 1.0
apiKeyFile: api_key
blocks:
- blockId: batch-1
pluginName: BatchGptPlugin
blockRuns: 5
configuration:
requestParams:
model: gpt-3.5-turbo
temperature: 0.7
top_p: 1
max_tokens: 500
executions:
- id: batch-1
dataFile: batch-data.json
prompt: batch.promptThe prompt is a simple Hello World prompt
Write me a paragraph about the world I live in
World: ```${my_world}```
The batch-data.json file contains the batch data.
{
"my_world" : [
"Hello World, I live in a magical place with magical creatures",
"Hello World, I live in a futuristic Utopia",
"Hello World, I live in a futuristic Dystopia"
]
}Modify these two files with your own data. For more information about batches
The following asks how many times to run the prompt request. If you choose 5 times, it will run all the prompt request 5 times.
? Number of Times to Run The Block: › 5
Next choose the output format. Do you just want a straight text response, or do you want it in JSON format.
? Response Format ›
JSON
❯ TEXT
If you choose JSON, you will be asked if you want to enable fixing the JSON response. This will attempt to parse any extraneous text that the AI assistant may add.
? Attempt to FIX JSON Responses? (y/n) › no
The project file will look like
---
projectName: prompt-1
projectVersion: 1.0
apiKeyFile: api_key
blocks:
- blockId: single-1
pluginName: ExperimentGptPlugin
blockRuns: 5
configuration:
requestParams:
model: gpt-3.5-turbo
temperature: 1.2
top_p: 1
max_tokens: 500
executions:
- id: exp-1
responseFormat: json
fixJson: false
promptChain:
- prompt-json.prompt
properties:
character: Commander in Starfleet
mainCharacterName: ''
- blockId: report-1
pluginName: ReportingGptPlugin
executions:
- id: report-1
blockIds:
- single-1The file includes some default values for the OpenAI requests. Change them to suit your needs. By default, it also adds the reporting plugin which generated HTML output of the user/assistant response.
The prompt-json.prompt looks like the following. Note how the output specifies to use JSON. Modify the prompt and properties to your needs.
Write me a story about ${character}. The main character is ${mainCharacterName}.
If no main character is given, choose one. Write one sentence only.
The response should be in JSON using the following structure:
Only use these fields. {"mainCharacterName": "", "story": ""}
For more information about prompts
Choose the Chain project archetype. Then go through the options.
? Number of Times to Run The Block: › 1
? Attempt to FIX JSON Responses? (y/n) › yes
? Number of Times to Run The Prompt Chain: › 2
The generated project.yaml file.
---
projectName: "chain-project"
projectVersion: "1.0"
apiKeyFile: "api_key"
blocks:
# Block demonstrates the use of importing properties
- blockId: chain-1
pluginName: ExperimentGptPlugin
blockRuns: 1 # Number of Stories
configuration:
requestParams:
model: gpt-3.5-turbo
temperature: 1.2
top_p: 1
max_tokens: 500
executions:
- id: exp-1-import
chainRuns: 2 # Number of times to run the promptChain
promptChain:
- story.prompt
- user-action.prompt # Simulates user input
excludesMessageHistory:
- user-action.prompt
fixJson: true
responseFormat: json
# Import properties from a properties file
import:
propertiesFile: properties.json # predefined values
properties:
planet: 1 # Earth
action: 3 # Lands on the planet
- blockId: report-1
pluginName: ReportingGptPlugin
executions:
- id: report-1
blockIds:
- chain-1The property fields in the above project.yaml file point to the index within the properties.json file is below. This file allows you to easily change test input.
{
"planet": [
"Earth",
"Venus",
"Jupiter"
],
"action": [
"Blows up the planet",
"Observes the planet From Orbit",
"Lands on the planet",
"Zips around the planet and hopes no one notices"
]
}The tool will substitute the planet "Earth" and the action "Lands on the planet" into the story prompt below. Notice that the AI will generate the character's name and the first paragraph of the story.
The response will be in JSON Format.
Captain ${captainsName} is near ${planet}. .
The last part of the story is: ${story}
Then the captain ${action}
Tell me a story about what happens next.
Be very descriptive. Write two sentences only.
Give me the captains name, if I haven't given it.
RESPONSE
The response must only be in JSON using the following structure.
Only use these fields. {"captainsName": "${captainsName}", "story": ""}
The tool will now pass the returned captain's name and the story from the first prompt into the user-action.prompt. We will get back an action that the character takes.
Give me an action for ${captainsName} for the following story:
${story}
The response must be in JSON using the following structure.
Only use these fields. {"action": ""}
Now we will run the story.prompt again but this time we will have both the captain's name and the next action he takes.
The follow is sample output from an actual run
As Captain John lands on the planet, he feels the trembling beneath his feet and sees the vibrant green flora around him.
He plants the Earth's flag to claim its new discovery and soon finds a thriving alien civilization welcoming him with open arms.
[user action "plants the flag to claim the new discovery"]
As Captain John plants the Earth's flag on the newfound planet, he is approached by the leaders of the alien civilization
who speak his language and reveal that they have known about Earth for centuries. They invite him to partake in a feast in
his honor, where he learns about their advanced technology and way of life.
Notice that chain run is the same as the number of paragraphs we have in the output. If we wanted another paragraph, we would set chainRuns to 3. If we had set blockRuns to 5, we would have generated 5 different stories.
For more information about chains
To generate an HTML report, add the ReportingGptPlugin as the last block. Under the blockIds add any previous block id that you want to add to the generated report.
---
projectName: experiment-reporting
projectVersion: '1.7'
apiKeyFile: "../../api_key"
blocks:
- blockId: chain-1
pluginName: ExperimentGptPlugin
blockRuns: 1
...
# Generate HTML Report
- blockId: report-1
pluginName: ReportingGptPlugin
executions:
- id: report-execution
blockIds:
- chain-1The report will display the entire chat for the configured block executions.
For more information about reporting
air --help
A command line tool for running GPT commands
Usage: air <command> [arguments]
Global options:
-h, --help Print this usage information.
Available commands:
clean Cleans project's output directory
count Returns the number of OpenApiCalls that would be made
genp Generates a new project
plugin Runs local version of ChatGPT Plugin
run Runs a project's blocks
Run "air help <command>" for more information about a command.
To clean a project, run the following
air clean
This deletes the output directory for the project.
Running OpenAI calls with a tool can be costly if you mis-configure it. To determine how many OpenAI calls a project will create, run the following command
air count
or for the count of a specific block
air count -b myblockId
It will output
Project: product-summary-2.8
Total OpenAPI Calls would be 12
If you want to know that your project is doing before incurring costs to OpenAI, use the dryRun flag.
air run --dryRun
Executing Block
Running Project: image-generation-2.3
BlockId: image-1, PluginName: ImageGptPlugin
----------
Starting Block Run: 1
Starting execution: 1 - Requires 1 calls to OpenAI
POST to https://api.openai.com/v1/images/generations
{"prompt":"Generate a picture of a Unicorn with a gold horn and wings","n":1,"size":"256x256","response_format":"url"}
Finished execution: 1
Starting execution: 2 - Requires 2 calls to OpenAI
POST to https://api.openai.com/v1/images/generations
{"prompt":"Generate a picture of a fish with giant eyes","n":1,"size":"256x256","response_format":"b64_json"}
POST to https://api.openai.com/v1/images/generations
{"prompt":"Generate a picture of a fish with giant eyes","n":1,"size":"512x512","response_format":"b64_json"}
Finished execution: 2
--------
Finished running project: 0 seconds