|
|
|
Show HN: GPTBanker – map/reduce over documents of any length with OpenAI
(wrotescan.com)
|
|
3 points
by aicharades
1236 days ago
|
|
Hi all! I created a app that can map/reduce over documents of an arbitrary length with OpenAI completions. You can enter in your OpenAI key and try it out. The backend is built with LangChain, which made it easy to manage the models, workflows (chains), and troubleshoot pipelines with buffer issues. Learnings:
- Tokens was the most helpful key for calls rather than characters or words. LangChain also just implemented a token splitter, which made it easier for me to adapt to documents of any type.
- Davinci performed much better than Curie for summarizing long documents over multiple chunks
- Oftentimes a map step was the only thing needed and the reduce/combine introduced more errors instead of improving context. |
|
Writer: Prompt: You are the writer Frank Herbert. Write a summary of a section of Dune based on the following input. The summary must be no longer than 650 characters long
Banker : Prompt: You are an investment banker. Write a company description in the style of a CIM based on a summary of the following section of an SEC filing. Focus on the company's operations. Do not use first person.
Storyteller : Prompt: You are a writer. Write a new story based on the section of a larger story you've been given. Include the character Equity Private and any employees of her private equity firm. Try to identify who Equity Private could be at the end of your story.