jiazhizhong/higress
1
0
Fork 0
mirror of https://github.com/alibaba/higress.git synced 2026-07-24 05:10:34 +08:00
Code Issues Packages Projects Releases Wiki Activity

fix(ai-proxy): fix cache token double-counting in Claude-to-OpenAI protocol conversion (#4149)

Browse Source 
Signed-off-by: Xijun Dai <daixijun1990@gmail.com>
Co-authored-by: 澄潭 <zty98751@alibaba-inc.com>
Co-authored-by: Kent Dong <ch3cho@qq.com>
This commit is contained in:
Xijun Dai
2026-07-18 00:14:24 +08:00
committed by GitHub
parent 165f2974db
commit 762324c376
2 changed files with 112 additions and 4 deletions
Download Patch File Download Diff File Expand all files Collapse all files

28
plugins/wasm-go/extensions/ai-proxy/provider/claude_to_openai.go
View File

@@ -353,6 +353,26 @@ func (c *ClaudeToOpenAIConverter) ConvertClaudeRequestToOpenAIWithOptions(body [
return result, nil
}
// computeClaudeInputTokens computes Claude-compatible input_tokens from OpenAI usage.
//
// In OpenAI's API, prompt_tokens includes cached_tokens (subset relationship).
// In Claude's API, input_tokens should NOT include cache tokens.
// We detect the OpenAI-standard semantics by checking if total_tokens == prompt_tokens + completion_tokens.
// For providers like Bedrock where prompt_tokens does NOT include cache tokens
// (total_tokens != prompt_tokens + completion_tokens), we return prompt_tokens as-is.
func computeClaudeInputTokens(u *usage) int {
if u == nil {
return 0
}
promptTokens := u.PromptTokens
if u.PromptTokensDetails != nil && u.PromptTokensDetails.CachedTokens > 0 {
if u.TotalTokens > 0 && u.TotalTokens == promptTokens+u.CompletionTokens {
return promptTokens - u.PromptTokensDetails.CachedTokens
}
}
return promptTokens
}
// ConvertOpenAIResponseToClaude converts an OpenAI response back to Claude format
func (c *ClaudeToOpenAIConverter) ConvertOpenAIResponseToClaude(ctx wrapper.HttpContext, body []byte) ([]byte, error) {
log.Debugf("[OpenAI->Claude] Original OpenAI response body: %s", string(body))
@@ -373,7 +393,7 @@ func (c *ClaudeToOpenAIConverter) ConvertOpenAIResponseToClaude(ctx wrapper.Http
// Only include usage if it's available
if openaiResponse.Usage != nil {
claudeResponse.Usage = claudeTextGenUsage{
InputTokens: openaiResponse.Usage.PromptTokens,
InputTokens: computeClaudeInputTokens(openaiResponse.Usage),
OutputTokens: openaiResponse.Usage.CompletionTokens,
}
if openaiResponse.Usage.PromptTokensDetails != nil {
@@ -582,7 +602,7 @@ func (c *ClaudeToOpenAIConverter) buildClaudeStreamResponse(ctx wrapper.HttpCont
// Only include usage if it's available
if openaiResponse.Usage != nil {
message.Usage = claudeTextGenUsage{
InputTokens: openaiResponse.Usage.PromptTokens,
InputTokens: computeClaudeInputTokens(openaiResponse.Usage),
OutputTokens: 0,
}
}
@@ -828,7 +848,7 @@ func (c *ClaudeToOpenAIConverter) buildClaudeStreamResponse(ctx wrapper.HttpCont
}
if openaiResponse.Usage != nil {
message.Usage = claudeTextGenUsage{
InputTokens: openaiResponse.Usage.PromptTokens,
InputTokens: computeClaudeInputTokens(openaiResponse.Usage),
OutputTokens: 0,
}
}
@@ -985,7 +1005,7 @@ func (c *ClaudeToOpenAIConverter) buildClaudeStreamResponse(ctx wrapper.HttpCont
openaiResponse.Usage.PromptTokens, openaiResponse.Usage.CompletionTokens)
usage := &claudeTextGenUsage{
InputTokens: openaiResponse.Usage.PromptTokens,
InputTokens: computeClaudeInputTokens(openaiResponse.Usage),
OutputTokens: openaiResponse.Usage.CompletionTokens,
}
if openaiResponse.Usage.PromptTokensDetails != nil {