Really enjoyed reading this! I like how you didn’t jump straight into the AI part—you built the recommender up step by step, from TF-IDF and collaborative filtering to embeddings and finally using an LLM to explain the recommendations.
The idea of separating the recommendation logic from the LLM explanation is especially nice. It makes the system feel more grounded and reduces the chance of the model making up information. Also, the “Milky Way” example is a great way to show why semantic search can be much more useful than simple keyword matching.
What I found most interesting is the focus on discovering less-famous parks rather than always recommending the usual ones. That makes the project feel genuinely useful, not just technically interesting. Great work!
Really enjoyed reading this! I like how you didn’t jump straight into the AI part—you built the recommender up step by step, from TF-IDF and collaborative filtering to embeddings and finally using an LLM to explain the recommendations.
The idea of separating the recommendation logic from the LLM explanation is especially nice. It makes the system feel more grounded and reduces the chance of the model making up information. Also, the “Milky Way” example is a great way to show why semantic search can be much more useful than simple keyword matching.
What I found most interesting is the focus on discovering less-famous parks rather than always recommending the usual ones. That makes the project feel genuinely useful, not just technically interesting. Great work!