text_rewritter

Text rewriter

Using LangChain and GPT3 to rewrite a large text chunk by chunk (remembering previous modifications to take them into account).

Theory

The text is sliced into chunks which are processed in order with the following prompt (shortened for the first two calls):

You are rewriting a novel, rewrite the full extract according to the instructions:

INSTRUCTIONS:
{instructions}

EXAMPLES OF REWRITINGS FROM OTHER PARTS OF THE TEXT:

{previous_rewrites}

EXTRACT TO REWRITE:
{previous_chunk}
{chunk}

REWRITTEN EXTRACT:
{previous_rewrite}

instructions is the user-provided prompt previous_rewrites are examples of rewrites whose chunk is similar to the provided chunk previous_chunk/previous_rewrite is the previous chunk of text (insuring continuity) chunk is the current chunk of text

Usage

• Source the mentalism_index conda env to get the dependencies
• Put the OpenAI API key in path
• put your input text and prompt in the data folder (it contains examples of prompts and results)
• Run rewrite.py.

I recommend testing your prompts on smaller texts as it can take many trials before finding a good prompt.

Potential improvements

• One could improve text size management to try and run on large text chunks while minimizing the risk of overflow (we could provide exactly as many examples as needed)
• let users add example rewritings to the search index in advance
• saving intermediate results to avoid loss and, overall, organize experiments
• add a best-of-3 interactive mode (better: we could let the user write their favorite version as an optional 4th option)