textjenerator 1.0.1

AI text generator

Text Jenerator

A Unified Python Interface for Modular Text Generation.

Text Jenerator is a framework designed to provide a consistent, abstract interface for deploying and interacting with various Large Language Models (LLMs). By separating the generic configuration and workflow from the model-specific implementation, it enables you to easily switch between different model backends—such as local GGUF models (via Llama-CPP), Hugging Face Transformers, or proprietary APIs (e.g., OpenAI, Gemini), with no code changes besides the config you pass.

Note that Text Jenerator is a library, not a chat app.

Features

• Modular Architecture: Uses Abstract Base Classes (ABC) to enforce a clean separation between the model details, meaning you only need to instantiate the TextGenerator class and call generate_text(), no matter which model you're using.
• Automatic Setup: Handles generic tasks like device detection (CPU/CUDA) and data type selection (bfloat16, float16, float32) in the base class.
• Plug-and-Play Models: Easily integrate new models by creating concrete child classes that implement two simple abstract methods (create_pipeline and run_pipeline).
• Model Registry: Uses a decorator (@register_model) to automatically track and load available generation backends.

Installation

pip install textjenerator

The core package has minimal dependencies. Model-specific implementations (e.g., for Llama-CPP or Transformers) require their respective libraries.

Models

Text Jenerator is BYOM. Currently textjenerator only supports llama-cpp-python on CPU only. So, you'll have to head over to Hugging Face and download a GGUF of your favourite model. As you'll be on CPU, you are probably best aiming for 1B to 7B parameter models, 4-bit to 8-bit quantised, depending on your system specs (higher is better but slower in both cases).

Add "model": "llama-cpp" to your config and the path to your GGUF file in the "model_path" attribute.

Quick Start & Usage

Config

Your project should include a config file (defaults are at textjenerator.config). A basic config might look like this:

config = {
    # model
    "model": "llama-cpp",
    "model_path": ".models/Llama-3.2-3B-Instruct-Q4_K_M.gguf",

    # hardware/system
    "device": "cpu",
    "dtype": "float32",
    "n_gpu_layers": -1,

    # LLM
    "max_context_size": 4096,
    "number_of_threads": 8 ,
    "verbose_warnings": False,
    "messages_to_keep_in_context": 4,
    "max_tokens_per_response": 256,
    "temperature": 0.7,
    "top_p": 0.9,
    "top_k": 50,
    "messages": [ 
          {"role": "system", "content": """You are Jenbot, an expert, helpful, and diligent assistant. You provide the user with accurate answers to their queries. You are polite, friendly, and a little sarcastic."""},
          {"role": "user", "content": """Hi, who are you?"""},
    ]
}

Run the Generation

To run the main workflow:

from textjenerator.models import registry
from config import config # or load your config however you prefer

text_generator = registry.get_model_class(config) # Choose which model you want from the registry
response = text_generator.generate_text() # create pipeline and run it
print(response)

You can also do these steps separately if you want to keep the model in memory, for example to have a back-and-forth conversation. In this case you can pass a new config, typically this will be to include the cumulative chat history, but you could also change other settings on the fly e.g.:

• Reduce temperature for a coding question
• Increase temperature for a creative writing request
• Increase max_tokens_per_response if you expect a longer reply

from textjenerator.models import registry
from config import config # or load your config however you prefer

text_generator = registry.get_model_class(config)
text_generator.create_pipeline()
response = text_generator.run_pipeline()
print(response)

# or pass another, doesn't have to be the whole thing:
new_config = {
    "messages": [ 
        {"role": "system", "content": """You are Jenbot, an expert, helpful, and diligent assistant. You provide the user with accurate answers to their queries. You are polite, friendly, and a little sarcastic."""},
        {"role": "user", "content": """Hi, who are you?"""},
        {"role": "assistant", "content": """You must be wondering who I am. Well, let me introduce myself: I'm Jenbot, your friendly AI assistant. I'm here to help answer any questions you may have, provide information on a wide range of topics, and even offer a dash of sarcasm when the situation calls for it."""},
        {"role": "user", "content": """What are your thoughts on the hard problem of consciousness?"""},
    ]
}
response = text_generator.run_pipeline(new_config)
print(response)

As you see here, to set up a chat with the bot, you have to pass the entire conversation history so far with each call to text_generator.run_pipeline(). With long chats, you might start getting erratic responses from smaller models (typically they start repeating themselves). Or you might hit memory errors. So you might want to trim your message list, e.g., keep only the most recent 8.

As you also see above, if you load the model and pass your config with registry.get_model_class(config), you don't need to pass the whole config each time you call text_generator.run_pipeline(new_config), only the bits you want to change.