store and serve language model prompts

prompt-serve helps you manage all of your large language model (LLM) prompts and associated settings/metadata in a straightforward, version controlled manner.

This project provides a YAML schema for storing prompts in a structured manner and a small API server that handles interactions with a Git repository, so you can treat prompts more like re-usable code.

• Release blog post

• YAML schema for prompts and associated metadata
• Associate prompts to one another to represent chains
• Create "packs" of multiple prompts or chains to represent categories of tasks or workflows
• Store any kind of prompt text or template
• Store LLM provider, model, and settings
• Easily convert to LangChain Prompt Templates
• Command-line utility for creating prompt files
• Command-line utility for viewing prompt statistics
• API server to upload or retrieve prompts*
• Version controlled via Git*

* = _coming soon_

Prompts follow the schema provided in schema.yml.

Check out the prompts repository to see it in action.

title: prompt-title-or-name
uuid: prompt-uuid
description: prompt-description
category: prompt-category
provider: model-provider
model: model-name
model_settings:
  temperature: 0.8
  top_k: 40
  top_p: 0.9
prompt: prompt-text
input_variables:
  - var1
  - var2
references:
  - https://example.com
  - https://example.com
associations:
  - prompt_uuid
  - prompt_uuid
packs:
  - pack-uuid
  - pack-uuid
tags:
  - tag
  - tag

You can use the validate.py utility to verify prompts meet the schema and have unique UUIDs.

By specifying the --create argument, a new UUID will be provided if a given prompt isn't unique for your scanned set. You can also gather statistics on the types of prompts in your collection by passing --gen-stats (see the next section for example stats output).

usage: validate.py [-h] [-s SCHEMA] [-f FILE] [-d DIRECTORY] [-c] [-g]

Validate YAML files against the prompt-serve schema.

options:
  -h, --help            show this help message and exit
  -s SCHEMA, --schema SCHEMA
                        schema file to validate against
  -f FILE, --file FILE  single file to validate
  -d DIRECTORY, --directory DIRECTORY
                        directory to validate
  -c, --create          create new uuids if validation fails
  -g, --gen-stats       generate statistics from directory

Example output

The command line utility stats.py will scan a directory of prompt-serve files and display statistics on the category, provider, and model fields in tables. Stats can also be optionally collected when running validate.py.

Example output

prompt-serve files can be easily converted to LangChain Prompt Templates.

The convert-to-langchain.py utility is provided to convert individual prompt-serve files to langchain. A simplified example of how to do this in Python is shown below.

import yaml
from langhain import PromptTemplate

def convert(path_to_ps_prompt):    
    with open(path_to_ps_prompt, 'r') as fp:
        data = yaml.safe_load(fp)
        prompt = data['prompt']
            
        if 'input_vars' in data.keys():
            input_vars = data['input_vars']
            langchain_template = PromptTemplate(template=prompt, input_variables=input_vars)
        else:
            langchain_template = PromptTemplate(template=prompt)
  
        return langchain_template

Example output

The command line utility create.py can be used to interactively create a prompt with the prompt-serve schema.

🪲 This is just a proof of concept and has a few known bugs:

• multi-line input for "prompt" field not handled correctly
• no defaults are set for optional fields

$ python create.py -n summary.yml
creating prompt file summary.yml ...
title (str): Summarize blog posts
description (str): Summarize a blog post with key takeaways
category (str): summarization
provider (str) : openai
model (str) : gpt-3.5-turbo
temperature (float) : 0.8
top_k (int) : 
top_p (float) : 0.9
max_tokens (int) : 512
stream (bool) : false
presence_penalty (float) : 
frequency_penalty (float) : 
prompt (str): Summarize the blog post provided below with 3-5 key takeaways as bullet points: {blog_content}
references (seq) : https://github.com/deadbits/prompt-serve
associations (seq) : 
packs (seq) : 
tags (seq) : 
 successfully wrote file summary.yml