## TODO About {#about} This is my first 'project' note page I'm trying. The idea is to put the entire process down as I ponder and work on things.

### The Idea {#the-idea} This is an idea I've had lolling about in the back of my head, one of those things that I'm pretty sure someone else has done, but I don't want to look it up out of fear of killing my motivation. I at least can guess some stupid SaaS company has made something similar, but probably tied into their own proprietary product. I'd much rather control the UI and the prompt myself.[^fn:1] I'm sure (most of you) know how much [spaced repetition software]({{< relref "spaced_repetition.md" >}}) rocks. If not check the underlying link for some resources. Anyways, software like anki can get complicated, but you can also essentially build your decks out of what are .csvs. And LLMs, with the proper guardrails are pretty darn good at assembling comma delimited files / JSON. You can kind of see where I'm going here, the overall idea is to use proper prompting to chain a topic out into subtopics, then sucessively build out a series of questions and answers on the topic, possibly using something like [Hong, Ruixin and Zhang, Hongming and Pan, Xiaoman and Yu, Dong and Zhang, Changshui :: Abstraction-of-Thought Makes Language Models Better Reasoners]({{< relref "../references/hongAbstractionofThoughtMakesLanguage2024.md" >}}) to "reason" about the questions and give less false positives. ```mermaid { exports="both" results="raw" } graph TD A[Start: Choose Topic 'SQL']:::Start -->|"Begin the process"| B[LLM: Generate High-Level Plan]:::LLM B -->|"Outline main areas"| C[LLM: Identify Key SQL Topics]:::LLM C -->|"Break down into subtopics"| D{For Each Topic}:::Decision D -->|"Topic 1"| E1[LLM: Generate Detailed Questions]:::LLM D -->|"Topic 2"| E2[LLM: Generate Detailed Questions]:::LLM D -->|"..."| E3[...]:::Process E1 --> F[Compile All Questions]:::Process E2 --> F E3 --> F F -->|"Review for quality"| G[LLM: Review and Refine Questions]:::LLM G -->|"Prepare for import"| H[Create CSV File / org-file structured for Anki]:::Process H -->|"Import to Anki"| I[Generate Anki Deck]:::Process I -->|"Ready for study"| J[Final Anki Deck on SQL]:::End %% Iterative loops G -->|"Insufficient quality"| C E1 -->|"Need more depth"| D E2 -->|"Need more depth"| D B -->|"Chain of thought"| C C -.->|"Chain of thought"| D E1 -.->|"Chain of thought"| F E2 -.->|"Chain of thought"| F G -.->|"Chain of thought"| H ``` High level I could possibly do this inside emacs itself, then keep up the generation as an artifact for people to wander into and maybe leave comments.[^fn:2] I'm not entirely sure if it would just be all babel-blocks or maybe a package that builds on something like [gptel.](https://github.com/karthink/gptel/blob/master/gptel-ollama.el) It's also worth pondering the usage of something like gpt4o-mini. While it's not free, it's cheap enough that I think the cost of an entire deck on a topic would be less than a dollar or two. I certainly spend money on dumber things. I could do this more easily in Python, I suppose, but I kind of wanted to tinker around with elisp more. Maybe use it to interact with the anki packages that exist.

### The Attempt {#the-attempt}

#### Making sure things are working {#making-sure-things-are-working} Of course, I'd like to iteratively drill down into actually what I want. It might be just the data scientist in me, but a bad habit I have is using notebooks / babel as a first run of things. Maybe it's a good thing, I dunno. First, a quick test of gptel, to make sure it's working. From the gptel-quick package, it seems like the request is meant to be the main point of entry for writing custom prompts. ```emacs-lisp (require 'gptel) (defun gptel-hello-world (name) (let ((result nil) (done nil)) (gptel-request (format "Say a friendly hello to %s in one short sentence." name) :callback (lambda (response _info) (setq result response done t))) (while (not done) (sleep-for 0.1)) result)) (gptel-hello-world "Justin") ```

Hello Justin, it's great to chat with you!

I got it working with ellama but I wanted to see if gptel works for chaining prompts. Seems like it does. I would note that for all my experiments I'm using a local small-ish version of llama 3. I'll probably switch to gpt4o-mini/turbo for the end. Maybe Claude? ```emacs-lisp (require 'gptel) (defun gptel-chain-of-thought (question) (let ((thoughts '()) (final-answer nil) (done nil)) ;; Initial thought (gptel-request (format "Consider this question: %s\nWhat's the first step in answering \ this? Provide your initial thought." question) :callback (lambda (response _info) (push response thoughts) (setq done t))) (while (not done) (sleep-for 0.1)) (setq done nil) ;; Generate a series of thoughts (dotimes (_ 3) ; Adjust the number of thoughts as needed (gptel-request (format "Based on your previous thought: %s\nWhat's the next step in \ your reasoning?" (car thoughts)) :callback (lambda (response _info) (push response thoughts) (setq done t))) (while (not done) (sleep-for 0.1)) (setq done nil)) ;; Final answer (gptel-request (format "Based on your chain of thoughts:\n%s\nWhat's your final answer \ to the original question: %s" (mapconcat 'identity (reverse thoughts) "\n") question) :callback (lambda (response _info) (setq final-answer response done t))) (while (not done) (sleep-for 0.1)) ;; Display the chain of thought (dolist (thought (reverse thoughts)) (message "Thought: %s" thought)) (message "Final Answer: %s" final-answer) final-answer)) ;; Example usage (gptel-chain-of-thought "I'm making an anki deck for SQL, what are 10 sub-topics to study") ```

Based on our conversation: 1. Data types (int, varchar, datetime) 2. Table operations (CREATE, ALTER, DROP) 3. Querying (SELECT, FROM, WHERE, JOIN) 4. Data manipulation (INSERT, UPDATE, DELETE) 5. Views and subqueries 6. Indexes and optimization techniques 7. SQL functions and aggregations (e.g., SUM, AVG, COUNT) 8. Common errors and pitfalls to avoid 9. Advanced SQL topics (e.g., window functions, common table expressions) 10. Best practices for writing efficient and readable SQL code

#### More Experiments with Chain/Analysis of Thought {#more-experiments-with-chain-analysis-of-thought} Now that we know it's simple to chain prompts in gptel, let's try using it for the initial part of my idea.[^fn:3] Maybe let's try it in Python just for ease..