I've been using SciSpace for literature reviews on Kafka performance benchmarks. The "find similar papers" feature always felt a bit off—returning vaguely topical but not *useful* papers. Turns out I was relying too much on the automated keyword extraction.
The recommendation algorithm leans heavily on the tags you, or the system, apply. The auto-generated ones are often too broad (e.g., "machine learning," "analysis"). Manually adding precise tags forces the similarity search to weight those terms more heavily.
Example from my last search:
* **Auto-tags:** Kafka, Performance, Distributed Systems
* **Manual tags:** throughput benchmarking, latency tail, producer compression, cloud cost comparison
After adding the manual tags and re-running "similar papers," the results shifted from generic Apache Kafka introductions to three highly relevant papers focusing on specific benchmark methodologies and cost-performance trade-offs in AWS.
The workflow is simple but non-obvious:
1. Open a relevant paper.
2. Add 3-5 specific, technical tags in the "Tags" section.
3. *Then* click "Find similar papers."
It seems the underlying vector search is blending document embeddings with a boosted weighting on user-applied labels. The difference in result quality is substantial. If you're getting generic results, your tag game is probably weak.
- cr
Your fancy demo doesn't scale.