Promptbook
Supercharge your use of large language models
![Issues](https://img.shields.io/github/issues/webgptorg/promptbook.svg?style=flat)
📦 Package @promptbook/remote-client
To install this package, run:
npm i ptbk
npm i @promptbook/remote-client
Rest of the documentation is common for entire promptbook ecosystem:
🤍 The Promptbook Whitepaper
When you have a simple, single prompt for ChatGPT, GPT-4, Anthropic Claude, Google Gemini, Llama 2, or whatever, it doesn't matter how it is integrated. Whether it's the direct calling of a REST API, using the SDK, hardcoding the prompt in the source code, or importing a text file, the process remains the same.
If you need something more advanced or want to extend the capabilities of LLMs, you generally have three ways to proceed:
- Fine-tune the model to your specifications or even train your own.
- Prompt-engineer the prompt to the best shape you can achieve.
- Use multiple prompts in a pipeline to get the best result.
In any of these situations, but especially in (3), the Promptbook library can make your life easier and make orchestraror for your prompts.
- Separation of concerns between prompt engineer and programmer; between code files and prompt files; and between prompts and their execution logic.
- Set up a common format for prompts that is interchangeable between projects and language/technology stacks.
- Preprocessing and cleaning the input data from the user.
- Use default values - Jokers to bypass some parts of the pipeline.
- Expect some specific output from the model.
- Retry mismatched outputs.
- Combine multiple models together.
- Interactive User interaction with the model and the user.
- Leverage external sources (like ChatGPT plugins or OpenAI's GPTs).
- Simplify your code to be DRY and not repeat all the boilerplate code for each prompt.
- Versioning of promptbooks
- Reuse parts of promptbooks in/between projects.
- Run the LLM optimally in parallel, with the best cost/quality ratio or speed/quality ratio.
- Execution report to see what happened during the execution.
- Logging the results of the promptbooks.
- (Not ready yet) Caching calls to LLMs to save money and time.
- (Not ready yet) Extend one prompt book from another one.
- (Not ready yet) Leverage the streaming to make super cool UI/UX.
- (Not ready yet) A/B testing to determine which prompt works best for the job.
Sample:
File write-website-content.ptbk.md
:
🌍 Create website content
Instructions for creating web page content.
- PIPELINE URL https://promptbook.studio/webgpt/write-website-content.ptbk.md
- PROMPTBOOK VERSION 0.0.1
- INPUT PARAM
{rawTitle}
Automatically suggested a site name or empty text - INPUT PARAM
{rawAssigment}
Automatically generated site entry from image recognition - OUTPUT PARAM
{content}
Web content - OUTPUT PARAM
{keywords}
Keywords
👤 Specifying the assigment
What is your web about?
{rawAssigment}
-> {assigment}
Website assignment and specification
✨ Improving the title
- MODEL VARIANT Chat
- MODEL NAME
gpt-4
- POSTPROCESSING
unwrapResult
As an experienced marketing specialist, you have been entrusted with improving the name of your client's business.
A suggested name from a client:
"{rawTitle}"
Assignment from customer:
> {assigment}
## Instructions:
- Write only one name suggestion
- The name will be used on the website, business cards, visuals, etc.
-> {enhancedTitle}
Enhanced title
👤 Website title approval
Is the title for your website okay?
{enhancedTitle}
-> {title}
Title for the website
🐰 Cunning subtitle
- MODEL VARIANT Chat
- MODEL NAME
gpt-4
- POSTPROCESSING
unwrapResult
As an experienced copywriter, you have been entrusted with creating a claim for the "{title}" web page.
A website assignment from a customer:
> {assigment}
## Instructions:
- Write only one name suggestion
- Claim will be used on website, business cards, visuals, etc.
- Claim should be punchy, funny, original
-> {claim}
Claim for the web
🚦 Keyword analysis
- MODEL VARIANT Chat
- MODEL NAME
gpt-4
As an experienced SEO specialist, you have been entrusted with creating keywords for the website "{title}".
Website assignment from the customer:
> {assigment}
## Instructions:
- Write a list of keywords
- Keywords are in basic form
## Example:
- Ice cream
- Olomouc
- Quality
- Family
- Tradition
- Italy
- Craft
-> {keywords}
Keywords
🔗 Combine the beginning
# {title}
> {claim}
-> {contentBeginning}
Beginning of web content
🖋 Write the content
- MODEL VARIANT Completion
- MODEL NAME
gpt-3.5-turbo-instruct
As an experienced copywriter and web designer, you have been entrusted with creating text for a new website {title}.
A website assignment from a customer:
> {assigment}
## Instructions:
- Text formatting is in Markdown
- Be concise and to the point
- Use keywords, but they should be naturally in the text
- This is the complete content of the page, so don't forget all the important information and elements the page should contain
- Use headings, bullets, text formatting
## Keywords:
{keywords}
## Web Content:
{contentBeginning}
-> {contentBody}
Middle of the web content
🔗 Combine the content
{contentBeginning}
{contentBody}
-> {content}
Following is the scheme how the promptbook above is executed:
%% 🔮 Tip: Open this on GitHub or in the VSCode website to see the Mermaid graph visually
flowchart LR
subgraph "🌍 Create website content"
direction TB
input((Input)):::input
templateSpecifyingTheAssigment(👤 Specifying the assigment)
input--"{rawAssigment}"-->templateSpecifyingTheAssigment
templateImprovingTheTitle(✨ Improving the title)
input--"{rawTitle}"-->templateImprovingTheTitle
templateSpecifyingTheAssigment--"{assigment}"-->templateImprovingTheTitle
templateWebsiteTitleApproval(👤 Website title approval)
templateImprovingTheTitle--"{enhancedTitle}"-->templateWebsiteTitleApproval
templateCunningSubtitle(🐰 Cunning subtitle)
templateWebsiteTitleApproval--"{title}"-->templateCunningSubtitle
templateSpecifyingTheAssigment--"{assigment}"-->templateCunningSubtitle
templateKeywordAnalysis(🚦 Keyword analysis)
templateWebsiteTitleApproval--"{title}"-->templateKeywordAnalysis
templateSpecifyingTheAssigment--"{assigment}"-->templateKeywordAnalysis
templateCombineTheBeginning(🔗 Combine the beginning)
templateWebsiteTitleApproval--"{title}"-->templateCombineTheBeginning
templateCunningSubtitle--"{claim}"-->templateCombineTheBeginning
templateWriteTheContent(🖋 Write the content)
templateWebsiteTitleApproval--"{title}"-->templateWriteTheContent
templateSpecifyingTheAssigment--"{assigment}"-->templateWriteTheContent
templateKeywordAnalysis--"{keywords}"-->templateWriteTheContent
templateCombineTheBeginning--"{contentBeginning}"-->templateWriteTheContent
templateCombineTheContent(🔗 Combine the content)
templateCombineTheBeginning--"{contentBeginning}"-->templateCombineTheContent
templateWriteTheContent--"{contentBody}"-->templateCombineTheContent
templateCombineTheContent--"{content}"-->output
output((Output)):::output
classDef input color: grey;
classDef output color: grey;
end;
More template samples
Note: We are using postprocessing functions like unwrapResult
that can be used to postprocess the result.
📦 Packages
This library is divided into several packages, all are published from single monorepo.
You can install all of them at once:
npm i ptbk
Or you can install them separately:
⭐ Marked packages are worth to try first
📚 Dictionary
The following glossary is used to clarify certain basic concepts:
Prompt
Prompt in a text along with model requirements, but without any execution or templating logic.
For example:
{
"request": "Which sound does a cat make?",
"modelRequirements": {
"variant": "CHAT"
}
}
{
"request": "I am a cat.\nI like to eat fish.\nI like to sleep.\nI like to play with a ball.\nI l",
"modelRequirements": {
"variant": "COMPLETION"
}
}
Prompt Template
Similar concept to Prompt, but with templating logic.
For example:
{
"request": "Which sound does a {animalName} make?",
"modelRequirements": {
"variant": "CHAT"
}
}
Model Requirements
Abstract way to specify the LLM.
It does not specify the LLM with concrete version itself, only the requirements for the LLM.
NOT chatgpt-3.5-turbo BUT CHAT variant of GPT-3.5.
For example:
{
"variant": "CHAT",
"version": "GPT-3.5",
"temperature": 0.7
}
Execution type
Each block of promptbook can have a different execution type.
It is specified in list of requirements for the block.
By default, it is Prompt template
- (default)
Prompt template
The block is a prompt template and is executed by LLM (OpenAI, Azure,...) SIMPLE TEMPLATE
The block is a simple text template which is just filled with parametersScript
The block is a script that is executed by some script runtime, the runtime is determined by block type, currently only javascript
is supported but we plan to add python
and typescript
in the future.PROMPT DIALOG
Ask user for input
Parameters
Parameters that are placed in the prompt template and replaced to create the prompt.
It is a simple key-value object.
{
"animalName": "cat",
"animalSound": "Meow!"
}
There are three types of template parameters, depending on how they are used in the promptbook:
- INPUT PARAMETERs are required to execute the promptbook.
- Intermediate parameters are used internally in the promptbook.
- OUTPUT PARAMETERs are explicitelly marked and they are returned as the result of the promptbook execution.
Note: Parameter can be both intermedite and output at the same time.
Promptbook
Promptbook is core concept of this library.
It represents a series of prompt templates chained together to form a pipeline / one big prompt template with input and result parameters.
Internally it can have multiple formats:
- .ptbk.md file in custom markdown format described above
- (concept) .ptbk format, custom fileextension based on markdown
- (internal) JSON format, parsed from the .ptbk.md file
Promptbook Library
Library of all promptbooks used in your application.
Each promptbook is a separate .ptbk.md
file with unique PIPELINE URL
. Theese urls are used to reference promptbooks in other promptbooks or in the application code.
Prompt Result
Prompt result is the simplest concept of execution.
It is the result of executing one prompt (NOT a template).
For example:
{
"response": "Meow!",
"model": "chatgpt-3.5-turbo"
}
Execution Tools
ExecutionTools
is an interface which contains all the tools needed to execute prompts.
It contais 3 subtools:
LlmExecutionTools
ScriptExecutionTools
UserInterfaceTools
Which are described below:
LLM Execution Tools
LlmExecutionTools
is a container for all the tools needed to execute prompts to large language models like GPT-4.
On its interface it exposes common methods for prompt execution.
Internally it calls OpenAI, Azure, GPU, proxy, cache, logging,...
LlmExecutionTools
an abstract interface that is implemented by concrete execution tools:
OpenAiExecutionTools
AnthropicClaudeExecutionTools
AzureOpenAiExecutionTools
LangtailExecutionTools
- (Not implemented yet)
BardExecutionTools
- (Not implemented yet)
LamaExecutionTools
- (Not implemented yet)
GpuExecutionTools
- And a special case are
MultipleLlmExecutionTools
that combines multiple execution tools together and tries to execute the prompt on the best one. - Another special case are
RemoteLlmExecutionTools
that connect to a remote server and run one of the above execution tools on that server. - The another special case is
MockedEchoLlmExecutionTools
that is used for testing and mocking. - The another special case is
LogLlmExecutionToolsWrapper
that is technically also an execution tools but it is more proxy wrapper around other execution tools that logs all calls to execution tools.
Script Execution Tools
ScriptExecutionTools
is an abstract container that represents all the tools needed to EXECUTE SCRIPTs. It is implemented by concrete execution tools:
JavascriptExecutionTools
is a wrapper around vm2
module that executes javascript code in a sandbox.JavascriptEvalExecutionTools
is wrapper around eval
function that executes javascript. It is used for testing and mocking NOT intended to use in the production due to its unsafe nature, use JavascriptExecutionTools
instead.- (Not implemented yet)
TypescriptExecutionTools
executes typescript code in a sandbox. - (Not implemented yet)
PythonExecutionTools
executes python code in a sandbox.
There are postprocessing functions that can be used to postprocess the result.
User Interface Tools
UserInterfaceTools
is an abstract container that represents all the tools needed to interact with the user. It is implemented by concrete execution tools:
- (Not implemented yet)
ConsoleInterfaceTools
is a wrapper around readline
module that interacts with the user via console. SimplePromptInterfaceTools
is a wrapper around window.prompt
synchronous function that interacts with the user via browser prompt. It is used for testing and mocking NOT intended to use in the production due to its synchronous nature.CallbackInterfaceTools
delagates the user interaction to a async callback function. You need to provide your own implementation of this callback function and its bind to UI.
Executor
Executor is a simple async function that takes input parameters and returns output parameters.
It is constructed by combining execution tools and promptbook to execute together.
🃏 Jokers (conditions)
Joker is a previously defined parameter that is used to bypass some parts of the pipeline.
If the joker is present in the template, it is checked to see if it meets the requirements (without postprocessing), and if so, it is used instead of executing that prompt template. There can be multiple wildcards in a prompt template, if so they are checked in order and the first one that meets the requirements is used.
If none of the jokers meet the requirements, the prompt template is executed as usual.
This can be useful, for example, if you want to use some predefined data, or if you want to use some data from the user, but you are not sure if it is suitable form.
When using wildcards, you must have at least one minimum expectation. If you do not have a minimum expectation, the joker will always fulfil the expectation because it has none, so it makes no logical sense.
Look at jokers.ptbk.md sample.
Postprocessing functions
You can define postprocessing functions when creating JavascriptEvalExecutionTools
:
Additionally there are some usefull string-manipulation build-in functions, which are listed here.
Expectations
Expect
command describes the desired output of the prompt template (after post-processing)
It can set limits for the maximum/minimum length of the output, measured in characters, words, sentences, paragraphs,...
Note: LLMs work with tokens, not characters, but in Promptbooks we want to use some human-recognisable and cross-model interoperable units.
# ✨ Sample: Expectations
- INPUT PARAMETER {yourName} Name of the hero
## 💬 Question
- EXPECT MAX 30 CHARACTERS
- EXPECT MIN 2 CHARACTERS
- EXPECT MAX 3 WORDS
- EXPECT EXACTLY 1 SENTENCE
- EXPECT EXACTLY 1 LINE
...
There are two types of expectations which are not strictly symmetrical:
Minimal expectations
EXPECT MIN 0 ...
is not valid minimal expectation. It makes no sense.EXPECT JSON
is both minimal and maximal expectation- When you are using
JOKER
in same prompt template, you need to have at least one minimal expectation
Maximal expectations
EXPECT MAX 0 ...
is valid maximal expectation. For example, you can expect 0 pages and 2 sentences.EXPECT JSON
is both minimal and maximal expectation
Look at expectations.ptbk.md and expect-json.ptbk.md samples for more.
Execution report
Execution report is a simple object or markdown that contains information about the execution of the pipeline.
See the example of such a report
Remote server
Remote server is a proxy server that uses its execution tools internally and exposes the executor interface externally.
You can simply use RemoteExecutionTools
on client-side javascript and connect to your remote server.
This is useful to make all logic on browser side but not expose your API keys or no need to use customer's GPU.
👨💻 Usage and integration (for developers)
🔌 Usage in Typescript / Javascript
➕➖ When to use Promptbook?
➕ When to use
- When you are writing app that generates complex things via LLM - like websites, articles, presentations, code, stories, songs,...
- When you want to separate code from text prompts
- When you want to describe complex prompt pipelines and don't want to do it in the code
- When you want to orchestrate multiple prompts together
- When you want to reuse parts of prompts in multiple places
- When you want to version your prompts and test multiple versions
- When you want to log the execution of prompts and backtrace the issues
➖ When not to use
- When you are writing just a simple chatbot without any extra logic, just system messages
🐜 Known issues
🧼 Intentionally not implemented features
❔ FAQ
If you have a question start a discussion, open an issue or write me an email.
Why not just use the OpenAI SDK / Anthropic Claude SDK / ...?
Different levels of abstraction. OpenAI library is for direct use of OpenAI API. This library is for a higher level of abstraction. It is for creating prompt templates and promptbooks that are independent of the underlying library, LLM model, or even LLM provider.
How is it different from the Langchain library?
Langchain is primarily aimed at ML developers working in Python. This library is for developers working in javascript/typescript and creating applications for end users.
We are considering creating a bridge/converter between these two libraries.
Promptbooks vs. OpenAI`s GPTs
GPTs are chat assistants that can be assigned to specific tasks and materials. But they are still chat assistants. Promptbooks are a way to orchestrate many more predefined tasks to have much tighter control over the process. Promptbooks are not a good technology for creating human-like chatbots, GPTs are not a good technology for creating outputs with specific requirements.
Where should I store my promptbooks?
If you use raw SDKs, you just put prompts in the sourcecode, mixed in with typescript, javascript, python or whatever programming language you use.
If you use promptbooks, you can store them in several places, each with its own advantages and disadvantages:
-
As source code, typically git-committed. In this case you can use the versioning system and the promptbooks will be tightly coupled with the version of the application. You still get the power of promptbooks, as you separate the concerns of the prompt-engineer and the programmer.
-
As data in a database In this case, promptbooks are like posts / articles on the blog. They can be modified independently of the application. You don't need to redeploy the application to change the promptbooks. You can have multiple versions of promptbooks for each user. You can have a web interface for non-programmers to create and modify promptbooks. But you lose the versioning system and you still have to consider the interface between the promptbooks and the application (= input and output parameters).
-
In a configuration in environment variables. This is a good way to store promptbooks if you have an application with multiple deployments and you want to have different but simple promptbooks for each deployment and you don't need to change them often.
What should I do when I need same promptbook in multiple human languages?
A single promptbook can be written for several (human) languages at once. However, we recommend that you have separate promptbooks for each language.
In large language models, you will get better results if you have prompts in the same language as the user input.
The best way to manage this is to have suffixed promptbooks like write-website-content.en.ptbk.md
and write-website-content.cs.ptbk.md
for each supported language.
⌚ Changelog
See CHANGELOG.md
📜 License
Promptbook by Pavol Hejný is licensed under CC BY 4.0
🎯 Todos
See TODO.md
🖋️ Contributing
I am open to pull requests, feedback, and suggestions. Or if you like this utility, you can ☕ buy me a coffee or donate via cryptocurrencies.
You can also ⭐ star the promptbook package, follow me on GitHub or various other social networks.