AI Paper Generator

AI app that generates scientific papers. It automatically searches for relevant literature and uses LLMs.

Features

• Generates all relevant parts of a paper: abstract, keywords, contents, conclusions, bibliography, etc.
• Utilizes LLM for writing text.
• Searches for relevant literature online.
• Program outputs all relevant information in JSON file, which can be used for programmatical processing.
• You can use Jinja tempates with generated contexts for making anything you want: simple text paper, LaTeX, Typst, or more!

How to Run this Project?

Store your OpenAI API key as an environmental variable:

export OPENAI_API_KEY=sk-******

THen, use uv:

uv run main.py

CLI Usage

Usage: main.py [OPTIONS] COMMAND [ARGS]...

Options:
  --help  Show this message and exit.

Commands:
  continue-checkpoint
  fill-template        Use Jinja templates to fill a context generated...
  generate-paper       Generate paper using the title.
  show-workflow-map    Visualize the internal workflow in images.

Paper generation:

Usage: main.py generate-paper [OPTIONS] TITLE LANGUAGE OUTPUT

  Generate paper using the title.

  Outputs JSON context.

Options:
  --templates-dir DIRECTORY  Directory for LLM message templates  [required]
  --llm-model TEXT           LLM model (currently, only OpenAI models are
                             supported)
  --embedding-model TEXT     LLM model (currently, only OpenAI models are
                             supported)
  --checkpoints-file PATH    Checkpoints file (will be overwritten).
  --help                     Show this message and exit.

Continue from checkpoint (in case generate-paper interrupted):

Usage: main.py continue-checkpoint [OPTIONS] [CHECKPOINTS_FILE] OUTPUT

Options:
  --templates-dir DIRECTORY  Directory for LLM message templates  [required]
  --llm-model TEXT           LLM model (currently, only OpenAI models are
                             supported)
  --embedding-model TEXT     LLM model (currently, only OpenAI models are
                             supported)
  --help                     Show this message and exit.

Fill Jinja template:

Usage: main.py fill-template [OPTIONS] CONTEXT TEMPLATE OUTPUT_PAPER
                             OUTPUT_BIBLIOGRAPHY

  Use Jinja templates to fill a context generated from `generate-paper`.

  `template` can be a directory of templates (root template must be
  `root.jinja`). It can be a file also.

Options:
  --help  Show this message and exit.

And, cherry on pie, visualize internal workflows:

Usage: main.py show-workflow-map [OPTIONS] OUTPUT_DIR

  Visualize the internal workflow in images.

Options:
  --help  Show this message and exit.

How this Project is Implemented?

Diagram of the most complex algorithm:

If you are thinking, what's intermediate mean, don't mind, this is some problem with LlamaIndex.

Tech stack:

• Python
• OpenAI API
• OpenAI's GPT and embedding models
• Semantic Scholar API (for paper search)
• LlamaIndex
• Especially - LlamaIndex Workflow