Model for keywording technical notes

Posted by olddoglearnsnewtrick@reddit | LocalLLaMA | View on Reddit | 4 comments

I've assembled a few thousand notes (stored in markdown text files) in the past years with "how to" on diverse subjects such as Linux/Mac managments, scripts, python, docker recipes, mongodb aggregation queries etc etc etc. Today I am searching through them with simple string matching (Craft app) and also I'd like to escape the flat dead simple subject classification I have now. Some of the notes I have under Debian are totally useful also in my MacOS folder etc etc. So I was thinking about using a LLM to ask and generate additional keywords to help me retrieve my recipes also with alternative searches. Is this a sensible idea? What approach do you suggest? Reading a given document say about installing a product, I would like to generate keywords such as linux, debian, macos, bash etc The notes are mostly in english, although my mothertongue sometimes slipped in (Italian) and what's most precious are the "codeblocks" which describe the actual code/steps I need to perform.

[-]

SomeOddCodeGuy@reddit

I do this exact thing, and I've had the best luck with Chain of Thought prompting and Hermes 2 Theta Llama 3 8b. I'll say that I chose that model for speed, but that I did try with bigger models and found that (obviously) bigger was better. But the most important thing for me was the Chain of Thought prompting. I'm going to give a very basic example of what that might look like here, but you really want to flesh out for yourself. Q: You are a very skilled keyword generation AI, and have been tasked with generating keywords based on an incoming prompt. Please see the prompt below: [ I have 2 cats. They like to play in boxes and attack strings. ] Please generate keywords that highlight the key nouns found within the above prompt. Please avoid generating keywords for verbs, adverbs, adjectives or pronouns. Please use only the singular tense for the keywords A: cat box string Q: You are a very skilled keyword generation AI, and have been tasked with generating keywords based on an incoming prompt. Please see the prompt below: [ There is a dog at the end of the road. It likes to chase cars and bark at squirrels. ] Please generate keywords that highlight the key nouns found within the above prompt. Please avoid generating keywords for verbs, adverbs, adjectives or pronouns. Please use only the singular tense for the keywords A: dog road car squirrel Q: You are a very skilled keyword generation AI, and have been tasked with generating keywords based on an incoming prompt. Please see the prompt below: [ {yourPromptGoesHere} ] Please generate keywords that highlight the key nouns found within the above prompt. Please avoid generating keywords for verbs, adverbs, adjectives or pronouns. Please use only the singular tense for the keywords A:

[-]

olddoglearnsnewtrick@reddit (OP)

Very interesting. Thanks a lot.

[-]

SomeOddCodeGuy@reddit

So I actually did the same thing lol. With that said, results were mixed, but I basically extended my CoT prompt to account for that. Here is an exact example that I'm using in my code. I have a few more CoT prompts before prompting the AI. Q: You are engaged in an ongoing conversation with a user. For each message, identify and extract key terms related to the topics discussed to help you recall relevant prior discussions from a text database. Conversation messages: [ Gerald: I really enjoy flying. It makes me happy I've always wanted to get a pilot's license. I think we talked about that before. I'd love to fly around in my own personal plane. Having it would be really helpful. Wonder if there's an app for getting more information about it? You wrote one didn't you? Gerald: Or should I get a boat? Ugh I have no idea. Ralf: ] Instructions: Analyze the message to pinpoint the main subjects. Select keywords that closely represent the specific topics discussed, focusing on proper nouns, technical terms, and specific products or tools mentioned. Avoid adjectives or sentiments that might introduce bias or trigger unrelated associations. For example, something like "app" could be also be "Application", "Program", or "Software". Please infer the user's meaning- if the user says "Thanks for finding that thing about prompting for. Reading that was helpful", you should include things they can read like "Paper", "Report", or "Article". A: Flying Pilot License Plane Airplane Aircraft App Application Program Code Coding Develop Boat Ship Seacraft

[-]

olddoglearnsnewtrick@reddit (OP)

Thank you very much for sharing

Model for keywording technical notes

Reply to Post

4 Comments

SomeOddCodeGuy@reddit

olddoglearnsnewtrick@reddit (OP)

SomeOddCodeGuy@reddit

olddoglearnsnewtrick@reddit (OP)