問題文
The same question returns different answers even with temperature set to zero. What accounts for the variation?
選択肢
- The retrieved passages differ between calls, so the model is given different context each time even though the generation settings are fixed
- The temperature setting does not apply to retrieval-based applications, since the passages are inserted after the sampling parameters have been evaluated
- The role assignment changes between calls because the token is refreshed, and the refreshed identity can read a different subset of the index
- The quota resets between calls and a request made just after the reset is processed with a different amount of available capacity