Skip to content

templates.md

Latest commit

 

History

History
233 lines (163 loc) · 6.2 KB

templates.md

File metadata and controls

233 lines (163 loc) · 6.2 KB

YAML Templates for llmquery

This document serves as a guide for creating YAML templates for llmquery. Templates are the foundation of how prompts are structured, validated, and rendered for various Language Model APIs.

Templates are powered by Jinja2, a Turing-complete template engine. This allows for the creation of dynamic and flexible templates through the use of conditional statements, loops, functions, and other advanced constructs.

Structure of a YAML Template

Each YAML template consists of the following key components:

1. Template ID

Every template must have a unique identifier specified under the id key.

id: unique-template-id

2. Template Metadata

Metadata allows users to add descriptive information about their templates. It provides context, notes, or categorization for better organization.

Common Metadata Fields

metadata:
    author: "Your Name"
    tags:
      - tag1
      - tag2
    category: "Your Category"
    description: "Your template description"

3. System Prompt

The system_prompt defines the behavior and tone of the LLM. This field describes the context and rules the model should follow.

system_prompt: >
    You are a helpful assistant.
  • Use > for multiline prompts.
  • Keep instructions concise and explicit.

4. Prompt

The prompt field contains the main query or instruction sent to the model. Use variables to make the prompt dynamic.

prompt: >
    Translate the following text from {{ source_language }} to {{ target_language }}:
    Text:
    {{ text }}

5. Variables

Variables allow the prompt to be dynamic and reusable. Define them under the variables key or pass them directly when instantiating the LLMQuery class. Variables provided in the LLMQuery class will overwrite those defined in the YAML template.

variables:
    source_language: English
    target_language: Spanish
    text: "Hello, how are you?"

Example of providing Variables in code:

from llmquery import LLMQuery

query = LLMQuery(
    provider="OPENAI",
    templates_path="./templates/translate-natural-language.yaml",
    variables={"source_language": "French", "target_language": "German", "text": "Bonjour"},
    openai_api_key="your-api-key"
)
response = query.Query()
print(response)

Example Templates

Here are some examples to help you create your own YAML templates.

Example 1: Language Detection

id: detect-natural-language

metadata:
    author: "Your Name"
    tags:
      - language-detection
    category: "Utility"

system_prompt: >
    You're an AI assistant. You should return the expected response without any additional information. The response should be exclusively in JSON with no additional code blocks or text.

prompt: >
    Analyze the following text and identify its language:
    Return response as: {"detected_language": "LANGUAGE_NAME"}
    {{ text }}

variables:
    text: " السلام عليكم"

Example 2: Find Book Author

id: find-book-author-name

metadata:
    author: "Your Name"
    tags:
      - book-info
    category: "Information Retrieval"

system_prompt: >
    You are an AI assistant that returns concise information in JSON format exclusively without any additional context, code blocks, or formatting.

prompt: >
    Who is the author of this book?
    Response should be in the format: {"author": "AUTHOR NAME"}
    Book name: {{ book }}

variables:
    book: Atomic Habits

Example 3: Translation

id: translate-natural-language

metadata:
    author: "Your Name"
    tags:
      - translation
      - language
    category: "Utility"

system_prompt: >
    You're an AI assistant. You should return the expected response without any additional information.

prompt: >
    Translate the following natural language from {{ source_language }} to {{ target_language }}:
    Text:

    {{ text }}

variables:
    source_language: English
    target_language: Spanish
    text: "Hello, how are you?"

Grammar and Best Practices

Grammar

  1. id: Unique string identifying the template.
  2. metadata: Optional descriptive information about the template.
  3. system_prompt: Explicit instructions for the LLM.
  4. prompt: The user instruction with variables embedded.
  5. variables: Key-value pairs defining dynamic components of the prompt.

Best Practices

  1. Clarity: Keep prompts clear and concise.
  2. JSON Responses: When applicable, enforce JSON-only outputs for structured responses.
  3. Error Handling: Provide explicit instructions in the system_prompt to handle errors gracefully.
  4. Dynamic Variables: Use variables to make the template reusable across different inputs.
  5. Validation: Test templates to ensure they produce the expected results.

Creating Your Own Template

  1. Start with the id section to uniquely identify your template.
  2. Add a metadata section to describe your template.
  3. Write a clear system_prompt to define the behavior of the model.
  4. Craft a prompt with placeholders for dynamic inputs.
  5. Define the variables with default values to test the template.

Example Template Skeleton

id: your-template-id

metadata:
    author: "Your Name"
    tags:
      - example-tag
    category: "Example Category"

system_prompt: >
    [Define the behavior of the LLM here.]

prompt: >
    [Write the main instruction or query here. Use {{ variable_name }} for placeholders.]

variables:
    variable_name: "default value"

Contribute Your Templates

We welcome contributions to the llmquery repository! If you've created useful YAML templates, consider submitting it to help the community.

Steps to Submit a Template

  1. Fork the repository from GitHub.
  2. Create a branch for your template addition.
  3. Add your template under the appropriate directory (e.g., templates/).
  4. Test your template to ensure it works as expected.
  5. Open a Pull Request (PR) with a clear description of your template and its use case.

Join us in building a library of powerful, reusable templates for everyone!

GitHub Link: https://github.com/mazen160/llmquery