@Comui520 is attempting to deploy a commit to the Confident AI Team on Vercel.

A member of the Team first needs to authorize it.

hey @Comui520 thanks for this. I believe the fix should be at the evaluate_chunk boundary instead however. The None is important because it does not mislead the user that the evaluation cost is indeed 0. Could you make that fix, and leave the other models alone? Thanks!

Done — moved the fix to the boundaries, left calculate_cost and all other models untouched. Here's what changed:

Files changed (3)

What the fix does

At every call site where cost returned by calculate_cost() is added to a running total, added a None guard:

# Before (crashes when pricing is unknown)
self.total_cost += cost

# After
if cost is not None:
    self.total_cost += cost

This prevents the TypeError that was silently swallowed by the broad except Exception in generate_contexts(), which previously caused empty goldens with no error surfaced.

The context generator and synthesizer now also raise explicit DeepEvalError instead of silently returning empty results when all documents fail or no contexts are produced.

Why this approach

• calculate_cost() returning None is the correct contract — "unknown price" ≠ "zero price"
• Guards at the boundary prevent arithmetic errors without misleading users about costs
• No changes to other model providers or tests

hey @Comui520 thanks for this. I believe the fix should be at the evaluate_chunk boundary instead however. The None is important because it does not mislead the user that the evaluation cost is indeed 0. Could you make that fix, and leave the other models alone? Thanks!嘿，谢谢。但我认为修复应该是在 evaluate_chunk 边界上，不过。 None 很重要，因为它不会误导用户认为评估成本确实是 0 。你能做这个修复，而让其他模型保持原样吗？谢谢！

@penguine-ip , appreciate your feedback. The None handling is now placed at the  evaluate_chunk  boundary, with other provider code untouched. Error visibility and new DeepSeek model support are also included.