How many questions realistically need to be saved to start increasing reuse percentage?

To start seeing a meaningful increase in your reuse percentage, you generally don't need a specific high volume, but rather high-density repetition.

In practice, you can begin to see improvements with as few as 50 to 100 well-curated examples, provided they represent the most common "bottleneck" queries your users ask.

Here is why that range is effective:

• The 80/20 Rule: In most systems, 80% of the traffic is driven by 20% of the questions. If you capture the top 20–30 recurring queries, your reuse rate will jump significantly because these are the questions your model is forced to solve from scratch most often.
• Quality over Quantity: A library of 1,000 messy or rarely asked questions will do less for your reuse percentage than 50 clean, canonical answers to high-frequency prompts.
• The Threshold of "Pattern Matching": Once you have enough saved responses to cover your most popular topics, the "hit rate" for your router increases. At the 50–100 mark, you have likely reached the point where a majority of routine user interactions can be intercepted by a saved response rather than requiring a full generative pass.

If you are just starting, focus on identifying the top 10 questions you see daily. Saving just those 10 can often account for a 10–20% increase in reuse efficiency immediately. Beyond that, the curve flattens, and you'll reach a point of diminishing returns where you need many more entries to capture the "long tail" of unique requests.