Interesting concept/research-project, but the results to just about every query I tried seem inaccurate and perplexing. Assuming the "similarity score" is meaningful, you may want to raise the cutoff or add an indicator (different color, fade, etc) for passages that get surfaced with a low match.