Algorithm Study Automation + alpha: A RAG-Based Similar Problem Recommendation System
Why Build This?
The Slack reminder bot I built around the Ebbinghaus forgetting curve was successful at making me revisit problems I had already solved. However, to keep the knowledge in long-term memory and actually improve algorithm skill, I need to solve variant problems or problems from the same category.
During review, I was manually giving the problem to an AI and asking, “Find problems similar to this.” Now I want to automate that step.
After about a week of solving and studying problems, the current pain points are:
Platform and language differences: After solving a Baekjoon problem, I have to ask an AI again to find similar problems from foreign platforms such as LeetCode or Codeforces.
Search noise: A similar title or statement does not necessarily mean the same algorithm is required. There can be many possible solution strategies, but for review I want to find problems that are as close as possible to the logic I practiced.
Finding Algorithm Problems
For algorithm problems, the data I need to match looks like this:
- Unnecessary parts (story): Story text attached to the problem, such as “Cheolsu buying apples.” This is search noise and should be removed.
- Algorithm form and tags: The core data structure and algorithmic logic, such as prefix sums on a two-dimensional array.
- Constraints: Time complexity and memory limits.
Technology: RAG-Based Architecture
Instead of simple keyword matching, I want to build a RAG (Retrieval-Augmented Generation) system that judges semantic similarity. I had been looking for a project where RAG would make sense, and this is a good opportunity.
The initial development flow is planned as follows.
1. Knowledge Base (Vector Database)
Collect LeetCode and Codeforces problem datasets and store them in a vector database such as Pinecone or ChromaDB. Instead of embedding the full problem statement, embed an algorithm summary preprocessed by an LLM to improve search accuracy.
2. Three-Step Pipeline
To improve matching quality, use the following process:
Metadata Filtering: First filter candidates using Baekjoon difficulty levels such as Gold and Silver, plus tags.
Semantic Search: Use OpenAI’s text-embedding-3-small model to compute similarity between problem structures.
LLM Re-Ranking: Send the top five candidates to GPT-4o mini and let it select the best first-ranked problem, including constraint compatibility.
Cost and Efficiency Analysis (Written with GPT)
Because this system runs inside the infrastructure of my personal blog, mainly GitHub Actions, it is expected to have almost no maintenance cost. Of course, I will only know after running it. There was some recent issue around GitHub Actions, but if Actions cannot be used, I can probably run it with Jenkins.
| Item | Technology | Cost for 30 Problems |
|---|---|---|
| Storage | Pinecone Starter Plan | $0 (Free) |
| Embedding | text-embedding-3-small | About $0.0006 |
| Inference/Verification | GPT-4o mini | About $0.15 |
| Compute | GitHub Actions | $0 (Public Repo) |
| Total | About $0.15, roughly 200 KRW |
Future Plan
Instead of only receiving a message that says “It is time to review,” I want an environment where Slack tells me, “Here is a LeetCode Medium problem similar to the Baekjoon problem you solved yesterday.”
Once this automation is complete, I will no longer need to copy a problem link and throw it at an AI just to find something similar.
One Slack message should be enough to continue studying. When the system is ready, I will share the concrete vector DB upload script and prompt engineering process.
댓글