github-featured / comet-ml--opik

发布

0 发布 8,340 Git 标签

RSS 订阅源
  • 1.9.34-3502 653ba0a99b
    比较

    [issue-4114] [BE] Add model name normalization for dot-based Claude model variants (#4254)

    frostbyte_neo 发布于 2025-12-01 10:06:15 +00:00

    • [issue-4114] [BE] Add model name normalization for dot-based Claude model variants

    Fixes cost estimation for Claude model names with dots (e.g., claude-3.5-sonnet)
    by normalizing them to hyphenated format (claude-3-5-sonnet) used in pricing database.

    Changes:

    • Add findModelPrice() method with backwards-compatible fallback logic
    • Add normalizeModelName() to replace dots with hyphens
    • Try exact match first (maintains backwards compatibility)
    • Fall back to normalized name if exact match fails
    • Add comprehensive tests verifying both normalization and backwards compatibility

    Resolves: #4114

    • Revision 2: Add test for versioned Claude models with dots (issue #4114)

    Adds specific test case for the exact models mentioned in the user complaint:

    • claude-sonnet-4.5-20250929
    • claude-haiku-4.5-20251001

    These versioned models now correctly normalize from dot notation
    to hyphenated format and return proper cost estimates.

    • Revision 3: Apply dot-to-hyphen normalization to ModelCapabilities and frontend

    Extends issue #4114 fix to ensure consistent model name normalization across:

    • Backend ModelCapabilities (vision capability detection)
    • Frontend modelCapabilities.ts (UI vision checks)

    This ensures vision capability detection works correctly for dot-notated models
    like "claude-3.5-sonnet", "gemini-1.5-pro", matching the same fix applied
    to cost calculation in CostService.

    Changes:

    • Update ModelCapabilities.normalize() to replace dots with hyphens
    • Update frontend normalizeModelName() with same logic
    • Add comprehensive tests for dot notation in ModelCapabilitiesTest
    • All 20 tests passing (11 ModelCapabilities + 9 CostService)
    • Revision 4: Address Copilot PR review comments

    Changes based on GitHub Copilot review feedback:

    1. Simplify normalizeModelName() method:

      • Remove redundant null/blank check since caller guarantees non-null
      • Document precondition in JavaDoc
      • Reduces unnecessary defensive code

    2. Rename test for clarity:

      • calculateCost_shouldHandleMultipleDotsInModelName_issue4114
        → calculateCost_shouldHandleUnknownModelWithDotsGracefully_issue4114
      • Test name now matches actual behavior (graceful handling of unknown models)
      • Updated assertion to check for exact zero (not >=)
      • Improved comments to explain what's being tested

    All 9 tests still passing.

    • Revision 5: Fix Prettier formatting in modelCapabilities.ts

    Remove trailing space in JSDoc comment that was causing linter error.
    All linting now passes.

    • Revision 2: Address code review comments
    • Remove @Nullable annotations from private method parameters per project convention
    • Use StringUtils.isBlank for more robust null/blank validation
    • Add lowercase normalization and case-insensitive comparison for better model matching
    • Consolidate CostServiceTest into parameterized test (15 test cases, was 9)
    • Consolidate ModelCapabilitiesTest into parameterized test (29 test cases, was 11)
    • Improves test maintainability while increasing coverage

    Changes per review feedback from @andrescrz and GitHub Copilot

    下载附件