Devlog Update: Fixing the Irrelevant Search Results Bug
Hey everyone,
As I mentioned in the last devlog, I ran into a significant bug where the search functionality was returning results for any query, completely ignoring whether the content was relevant to space-related themes.
The Fix
I’ve now resolved the core issue by switching to semantic search. Instead of relying on simple keyword or basic matching, the system now uses embeddings to understand the meaning and context of both the query and the stored data. This ensures that only genuinely relevant space-themed results are returned.
For the implementation, I integrated @xenova/transformers (the browser/Node.js friendly Transformers.js library). It allows to run powerful embedding models locally, which is great for privacy and offline capability.
Remaining Challenge
The main downside right now is that the semantic search is quite data-heavy. It’s putting significant load on memory, especially when handling larger datasets or multiple simultaneous queries. I’ll need to explore optimizations such as:

• Model quantization
• Smaller/faster embedding models
• Caching strategies
• Possible backend offloading
  Even with the memory issue still pending, this change fixes a major part of the bug and represents a big step forward in search quality.
  Thanks for following along more updates soon as I tackle the performance side of things!