Skip to content

Latest commit

 

History

History
741 lines (595 loc) · 13.2 KB

README.md

File metadata and controls

741 lines (595 loc) · 13.2 KB

Mext

Mext is a powerful text template language designed for crafting prompts for LLM.

The primary objective of Mext is to provide human friendly yet powerful prompt templates for LLM (large language model), emphasizing readability and adaptability.

At the same time, the versatility of Mext extends its utility beyond, making it a powerful tool for a wide array of text manipulation tasks.

This README is written with Mext. Source file: readme_src/README.mext.

Early Stage

The Mext language is still in its early stage. New syntaxes may be introduced and broken changes might be made.

Table of content

Installation

Use the following command to install mext.

$ pip install mext-lang

Development

To build and install mext from scratch, clone this repository and use make.

The default target of make will perform testing, cleanup previous build, and install Mext.

$ make

To perform tests, use:

$ make test

You may as well use pip to install an editable version directly:

$ pip install -e .

Render Mext file

You can render a Mext file handly with mext/scripts/render_mext.py.

Usage:

$ python -m mext.scripts.render_mext readme_src/README.mext -o README.md

Or, if you installed mext, the render script will be installed automatically:

$ render-mext readme_src/README.mext -o README.md

To render this README.md with make:

$ make README.md

This will also render the data file readme_src/README.yaml from readme_src/README-yaml.mext and readme_src/README-yaml.yaml.

Usage as a template language

Check out the syntax of Mext as well: Syntax

Basic usage

To compose a template with Mext, use Mext.compose.

Python:

from mext import Mext

mext = Mext()
prompt = mext.compose(template="""
This is an example template.
{@if not undefined name}
The name is: {name}
{@endif}
""", name="Yamato")
print(prompt)

Output:

This is an example template.
The name is: Yamato

Reuse template

You can set a template and use different variables with it.

Given a tempalte file reuse.mext:

The name is: {@if not novalue name}{name}{@else}Unknown{@endif}

Python:

from mext import Mext

mext = Mext()
mext.set_template(template_fn="reuse.mext")

prompt = mext.compose()
print(prompt)

prompt = mext.compose(name="Sydney")
print(prompt)

Output:

The name is: Unknown
The name is: Sydney

Use with statement

You can use use_template and use_params with the with statement.

Given a tempalte file with_template.mext:

This is the profile of {name}:
name: {name}
title: {title}
backend: {backend}

Python:

from mext import Mext

mext = Mext()
with mext.use_template(template_fn="with_template.mext"): # set the template
  with mext.use_params(name="Mike", title="AI Software Engineer"): # set the parameters for the template
    prompt = mext.compose(backend="Claude 3")
    print(prompt)

Output:

This is the profile of Mike:
name: Mike
title: AI Software Engineer
backend: Claude 3

Use custom formatters

register_formatter is a powerful function that enables custom functions for text processing.

Python:

from mext import Mext, MextParser
from urllib.parse import quote

parser = MextParser()
parser.register_formatter('encode_uri_component', quote) # register a custom formatter 'encode_uri_component'

mext = Mext()
mext.set_parser(parser)
prompt = mext.compose(template="""
{@format encode_uri_component var}
""", params={
  'var': "there are some spaces in this sentence",
})
print(prompt)

Output:

there%20are%20some%20spaces%20in%20this%20sentence

Render prompts for LLM

Mext can be integrated into existing LLM pipeline easily.

Integration with Langchain

Mext template:

Answer the following questions one by one:
{@for q in questions}
- {q}
{@endfor}

Answer:

Python:

from mext import Mext

from langchain_openai import ChatOpenAI
from langchain_core.runnables import RunnableLambda
from langchain_core.output_parsers import StrOutputParser

def questions2prompt(params):
  mext = Mext()
  prompt = mext.compose(template_fn="examples/prompts_for_llm.mext", params=params)
  print('==== PROMPT ====')
  print(prompt) # print the prompt for debug purpose
  return prompt

# Set your api key in environment: OPENAI_API_KEY=<your_api_key>
llm = ChatOpenAI(model_name="gpt-3.5-turbo")
output_parser = StrOutputParser()
chain = RunnableLambda(questions2prompt) | llm | output_parser

params = {
  "questions": [
    'The distance of sun to earth',
    'Who is Alan Turing',
    'What is the name of the planet where the Foundation established?',
  ],
}
output = chain.invoke(params)
print('==== OUTPUT ====')
print(output)

Example output:

==== PROMPT ====
Answer the following questions one by one:
- The distance of sun to earth
- Who is Alan Turing
- What is the name of the planet where the Foundation established?

Answer:
==== OUTPUT ====
1. The distance of the sun to earth is approximately 93 million miles.
2. Alan Turing was a British mathematician, logician, and computer scientist who is considered one of the founding fathers of computer science.
3. The name of the planet where the Foundation established is Terminus.

You can run this example with:

$ python examples/prompts_for_llm.py

Syntax

Note although the @import syntax is used in most of the examples in this section, in production it is more often that variables are passed to Mext.compose as parameters directly. Check out the section Usage as a template language as well.

if

Template:

{@import "if.yaml"}

{@if true}
This will be shown.
{@elif false}
This will not be shown.
{@else}
This will not be shown, either.
{@endif}

{@if not empty var}
{var}
{@endif}

{@if not undefined var}
var is defined.
{@endif}
{@if undefined var2}
var2 is undefined.
{@endif}

{@if novalue var2}
var2 has no value.
{@endif}

Given params:

{
  "var": "Text from variable."
}

Produce:

This will be shown.

Text from variable.

var is defined.
var2 is undefined.

var2 has no value.

for

Template:

{@import "for.yaml"}

## Array
{@for item in arr}
- name: {item[name]}
  content: {item[content]}
{@endfor}

## Dictionary
{@for item_key, item_val in dict}
- key: {item_key}
  val: {item_val}
{@endfor}

Given params:

{
  "arr": [
    {
      "name": "Item 1",
      "content": "Content 1"
    },
    {
      "name": "Item 2",
      "content": "Content 2"
    }
  ],
  "dict": {
    "key1": "Value 1",
    "key2": "Value 2"
  }
}

Produce:

## Array
- name: Item 1
  content: Content 1
- name: Item 2
  content: Content 2

## Dictionary
- key: key1
  val: Value 1
- key: key2
  val: Value 2

trim_newline

Template:

@trim_newline if the following next blocks produce empty result until it meets a non-empty block.
{@trim_newline}
{@if false}
This will be ignored.
{@endif}

{@trim_newline}
{@if true}
This will be shown right after the line above.
{@endif}

{@trim_newline}
The above and the next new line will not be trimed.

The end.

Produce:

@trim_newline if the following next blocks produce empty result until it meets a non-empty block.
This will be shown right after the line above.

The above and the next new line will not be trimed.

The end.

format

Template:

{@import "format.yaml"}
@format a variable using given format. Formatters can be registered with `parser.register_formatter`.
{@format json var}

Given params:

{
  "var": {
    "abstract": "List of fruits to purchase.",
    "fruits": [
      "apple",
      "banana",
      "pear"
    ]
  }
}

Produce:

@format a variable using given format. Formatters can be registered with `parser.register_formatter`.
{
  "abstract": "List of fruits to purchase.",
  "fruits": [
    "apple",
    "banana",
    "pear"
  ]
}

import

Template:

@import variables from a file.
File ends with 'yaml' or 'json' will automatically be loaded as an object. When 'as' clause is present, the object will be loaded as a variable.
```
{@import "import.yaml" as imported}
{@format json imported}
```

If no 'as' clause is given, first level members of the object will be loaded into the local namespace.
```
{@import "import.yaml"}
{imported_var}
```

Other file type will be loaded as a string variable. Note you must specify the 'as' clause in this case.
Loaded from "default.mext":
```
{@import "default.mext" as default_mext}
{default_mext}
```

Using variable as filename is also possible.
Loaded from "set.mext":
```
{@import imported.set_mext_fn as set_mext}
{set_mext}
```

Given params:

{
  "imported_var": "This variable is imported from a yaml file.",
  "set_mext_fn": "set.mext"
}

Produce:

@import variables from a file.
File ends with 'yaml' or 'json' will automatically be loaded as an object. When 'as' clause is present, the object will be loaded as a variable.
```
{
  "imported_var": "This variable is imported from a yaml file.",
  "set_mext_fn": "set.mext"
}
```

If no 'as' clause is given, first level members of the object will be loaded into the local namespace.
```
This variable is imported from a yaml file.
```

Other file type will be loaded as a string variable. Note you must specify the 'as' clause in this case.
Loaded from "default.mext":
```
@default will not overwrite the variable.
{@default var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}

When the variable exists:
{@import "default.yaml"}
{@default var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}
```

Using variable as filename is also possible.
Loaded from "set.mext":
```
{@import "set.yaml"}
@set will overwrite the variable.
{@set var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}
```

include

Template:

@include content from "default.mext" file:
```
{@include "default.mext"}
```

With parameters:
```
{@include "default.mext" var=true}
```

Produce:

@include content from "default.mext" file:
```
@default will not overwrite the variable.
var is false.

When the variable exists:
var is true.
```

With parameters:
```
@default will not overwrite the variable.
var is true.

When the variable exists:
var is true.
```

input

Template:

@input set a variable by calling a provided callback.
{@input var}

Use the `callbacks` parameter to definite a callback dictionary when calling `parser.parse`.
```
parser.parse(
  template=template,
  template_fn=template_fn,
  callbacks={{
    'var': lambda x: 'any function',
  }},
)
```

Produce:

@input set a variable by calling a provided callback.
any function

Use the `callbacks` parameter to definite a callback dictionary when calling `parser.parse`.
```
parser.parse(
  template=template,
  template_fn=template_fn,
  callbacks={
    'var': lambda x: 'any function',
  },
)
```

option

Template:

@option controls the behavior of the parser.
{@option final_strip off}
This will keep the empty line below.

Produce:

@option controls the behavior of the parser.
This will keep the empty line below.

set

Template:

{@import "set.yaml"}
@set will overwrite the variable.
{@set var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}

Given params:

{
  "var": true
}

Produce:

@set will overwrite the variable.
var is false.

default

Template:

@default will not overwrite the variable.
{@default var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}

When the variable exists:
{@import "default.yaml"}
{@default var false}
{@if var}
var is true.
{@else}
var is false.
{@endif}

Given params:

{
  "var": true
}

Produce:

@default will not overwrite the variable.
var is false.

When the variable exists:
var is true.

count

Template:

{@import "count.yaml"}

{@set idx 0}
{@for item in items}
{@count idx}
{idx}. {item}
{@endfor}

Given params:

{
  "items": [
    "Item 1",
    "Item 2",
    "Item 3"
  ]
}

Produce:

1. Item 1
2. Item 2
3. Item 3

comment

Template:

@comment will not be shown in the result.
{@comment}
This will be ignored.
{@endcomment}

Produce:

@comment will not be shown in the result.