Welcome to the Power Users community on Codidact!
Power Users is a Q&A site for questions about the usage of computer software and hardware. We are still a small site and would like to grow, so please consider joining our community. We are looking forward to your questions and answers; they are the building blocks of a repository of knowledge we are building together.
Post History
LLMs (AI) have limited context windows which makes it challenging to include all necessary context in the prompt. A common solution is a memory book. The memory book stores arbitrarily large amoun...
#2: Post edited
by
matthewsnyder
·
2024-03-13T20:45:04Z (about 2 months ago)
- LLMs (AI) have limited context windows which makes it challenging to include all necessary context in the prompt. A common solution is a memory book.
The memory book stores arbitrarily large amounts of text (typically past conversations as well as reference material like wikipedia articles), segmented in some way (sentence, paragraph, etc). When you type in a prompt, it uses a dumb text similarity metric (like cosine similarity) to retrieve the most relevant snippets and attaches them to your prompt for the AI's benefit. Cosine similarity has low precision but decent recall, so while not all passages it finds will be relevant, if there is something relevant it will end up as a hit, and the AI can sort them out on its own when doing more sophisticated processing of the prompt.This sounds like a very useful thing when writing notes, even without an AI. Is there any program intended for human use (GUI or CLI) which can scan my notes, build a database, and retrieve relevant passages easily based on some rough text similarity measure? (not just keywords)- Inspired by https://rpg.codidact.com/posts/285850
- LLMs (AI) have limited context windows which makes it challenging to include all necessary context in the prompt. A common solution is a memory book.
- The memory book stores arbitrarily large amounts of text (typically past conversations as well as reference material like wikipedia articles), segmented in some way (sentence, paragraph, etc). When you type in a prompt, it uses a dumb text similarity metric (like cosine similarity) to retrieve the most relevant snippets and attaches them to your prompt for the AI's benefit. Cosine similarity has low precision but decent recall, so while not all passages it finds will be relevant, if there is something relevant it will end up as a hit, and the AI can sort them out on its own when doing more sophisticated processing of the prompt.
- This sounds like a very useful thing when writing notes, even without an AI. Is there any program intended for human use (GUI or CLI) which can scan my notes, build a database, and retrieve relevant passages easily based on some rough text similarity measure? (not just keywords - eg. cosine can deal with synonyms)
- Inspired by https://rpg.codidact.com/posts/285850
#1: Initial revision
by
matthewsnyder
·
2024-03-13T20:41:00Z (about 2 months ago)
Is there a memory book for humans?
LLMs (AI) have limited context windows which makes it challenging to include all necessary context in the prompt. A common solution is a memory book. The memory book stores arbitrarily large amounts of text (typically past conversations as well as reference material like wikipedia articles), segmented in some way (sentence, paragraph, etc). When you type in a prompt, it uses a dumb text similarity metric (like cosine similarity) to retrieve the most relevant snippets and attaches them to your prompt for the AI's benefit. Cosine similarity has low precision but decent recall, so while not all passages it finds will be relevant, if there is something relevant it will end up as a hit, and the AI can sort them out on its own when doing more sophisticated processing of the prompt. This sounds like a very useful thing when writing notes, even without an AI. Is there any program intended for human use (GUI or CLI) which can scan my notes, build a database, and retrieve relevant passages easily based on some rough text similarity measure? (not just keywords) Inspired by https://rpg.codidact.com/posts/285850