diff --git a/internal/transformer/convert/claude_gemini.go b/internal/transformer/convert/claude_gemini.go index a7fc3c4..37d02fc 100644 --- a/internal/transformer/convert/claude_gemini.go +++ b/internal/transformer/convert/claude_gemini.go @@ -53,8 +53,17 @@ func ClaudeReqToGemini(claudeReq []byte, model string) ([]byte, error) { if req.MaxTokens > 0 { genConfig["maxOutputTokens"] = req.MaxTokens } - if req.Temperature > 0 { - genConfig["temperature"] = req.Temperature + if req.Temperature != nil { + genConfig["temperature"] = *req.Temperature + } + if req.TopP != nil { + genConfig["topP"] = *req.TopP + } + if req.TopK != nil { + genConfig["topK"] = *req.TopK + } + if len(req.StopSequences) > 0 { + genConfig["stopSequences"] = req.StopSequences } if len(genConfig) > 0 { geminiReq["generationConfig"] = genConfig diff --git a/internal/transformer/convert/claude_openai.go b/internal/transformer/convert/claude_openai.go index d9b2391..0ee53e7 100644 --- a/internal/transformer/convert/claude_openai.go +++ b/internal/transformer/convert/claude_openai.go @@ -116,9 +116,9 @@ func ClaudeReqToOpenAI(claudeReq []byte, model string) ([]byte, error) { if req.MaxTokens > 0 { openaiReq.MaxCompletionTokens = req.MaxTokens } - if req.Temperature > 0 { - openaiReq.Temperature = &req.Temperature - } + openaiReq.Temperature = req.Temperature + openaiReq.TopP = req.TopP + openaiReq.Stop = req.StopSequences // Convert tools if len(req.Tools) > 0 { diff --git a/internal/transformer/convert/claude_sampling_test.go b/internal/transformer/convert/claude_sampling_test.go new file mode 100644 index 0000000..cd8e1d1 --- /dev/null +++ b/internal/transformer/convert/claude_sampling_test.go @@ -0,0 +1,97 @@ +package convert + +import ( + "encoding/json" + "testing" +) + +const claudeReqWithSampling = `{ + "model": "claude-sonnet-4-6", + "max_tokens": 1024, + "temperature": 0, + "top_p": 0.9, + "top_k": 40, + "stop_sequences": ["END", "STOP"], + "messages": [{"role": "user", "content": "hi"}] +}` + +const claudeReqWithoutSampling = `{ + "model": "claude-sonnet-4-6", + "max_tokens": 1024, + "messages": [{"role": "user", "content": "hi"}] +}` + +func convertedBody(t *testing.T, out []byte, err error) map[string]interface{} { + t.Helper() + if err != nil { + t.Fatalf("conversion failed: %v", err) + } + var body map[string]interface{} + if err := json.Unmarshal(out, &body); err != nil { + t.Fatalf("converted request is not JSON: %v", err) + } + return body +} + +func TestClaudeReqToOpenAIKeepsSamplingParams(t *testing.T) { + out, err := ClaudeReqToOpenAI([]byte(claudeReqWithSampling), "gpt-test") + body := convertedBody(t, out, err) + + // temperature 0 is a real value (deterministic output), not "unset" + if v, ok := body["temperature"]; !ok || v != float64(0) { + t.Errorf("expected temperature 0 to be kept, got %v (present=%v)", v, ok) + } + if body["top_p"] != 0.9 { + t.Errorf("expected top_p 0.9, got %v", body["top_p"]) + } + stop, _ := body["stop"].([]interface{}) + if len(stop) != 2 || stop[0] != "END" || stop[1] != "STOP" { + t.Errorf("expected stop [END STOP], got %v", body["stop"]) + } + if _, ok := body["top_k"]; ok { + t.Errorf("OpenAI has no top_k, it should not be forwarded") + } +} + +func TestClaudeReqToOpenAIOmitsUnsetSamplingParams(t *testing.T) { + out, err := ClaudeReqToOpenAI([]byte(claudeReqWithoutSampling), "gpt-test") + body := convertedBody(t, out, err) + + for _, key := range []string{"temperature", "top_p", "stop"} { + if v, ok := body[key]; ok { + t.Errorf("expected %s to be omitted when the client did not set it, got %v", key, v) + } + } +} + +func TestClaudeReqToGeminiKeepsSamplingParams(t *testing.T) { + out, err := ClaudeReqToGemini([]byte(claudeReqWithSampling), "gemini-test") + body := convertedBody(t, out, err) + + cfg, _ := body["generationConfig"].(map[string]interface{}) + if v, ok := cfg["temperature"]; !ok || v != float64(0) { + t.Errorf("expected temperature 0 to be kept, got %v (present=%v)", v, ok) + } + if cfg["topP"] != 0.9 { + t.Errorf("expected topP 0.9, got %v", cfg["topP"]) + } + if cfg["topK"] != float64(40) { + t.Errorf("expected topK 40, got %v", cfg["topK"]) + } + stop, _ := cfg["stopSequences"].([]interface{}) + if len(stop) != 2 || stop[0] != "END" || stop[1] != "STOP" { + t.Errorf("expected stopSequences [END STOP], got %v", cfg["stopSequences"]) + } +} + +func TestClaudeReqToGeminiOmitsUnsetSamplingParams(t *testing.T) { + out, err := ClaudeReqToGemini([]byte(claudeReqWithoutSampling), "gemini-test") + body := convertedBody(t, out, err) + + cfg, _ := body["generationConfig"].(map[string]interface{}) + for _, key := range []string{"temperature", "topP", "topK", "stopSequences"} { + if v, ok := cfg[key]; ok { + t.Errorf("expected %s to be omitted when the client did not set it, got %v", key, v) + } + } +} diff --git a/internal/transformer/types.go b/internal/transformer/types.go index b4712e4..514f092 100644 --- a/internal/transformer/types.go +++ b/internal/transformer/types.go @@ -38,6 +38,8 @@ type OpenAIRequest struct { MaxTokens int `json:"max_tokens,omitempty"` // Legacy field MaxCompletionTokens int `json:"max_completion_tokens,omitempty"` Temperature *float64 `json:"temperature,omitempty"` + TopP *float64 `json:"top_p,omitempty"` + Stop []string `json:"stop,omitempty"` Stream bool `json:"stream,omitempty"` StreamOptions *StreamOptions `json:"stream_options,omitempty"` EnableThinking bool `json:"enable_thinking,omitempty"` // For models that support reasoning/thinking @@ -106,15 +108,20 @@ type ClaudeMessage struct { // ClaudeRequest represents a Claude API request type ClaudeRequest struct { - Model string `json:"model"` - Messages []ClaudeMessage `json:"messages"` - MaxTokens int `json:"max_tokens,omitempty"` - Temperature float64 `json:"temperature,omitempty"` - Stream bool `json:"stream,omitempty"` - System interface{} `json:"system,omitempty"` // Can be string or array of system messages - Thinking interface{} `json:"thinking,omitempty"` // Claude's thinking/extended thinking parameter - Tools []ClaudeTool `json:"tools,omitempty"` - ToolChoice interface{} `json:"tool_choice,omitempty"` + Model string `json:"model"` + Messages []ClaudeMessage `json:"messages"` + MaxTokens int `json:"max_tokens,omitempty"` + // Sampling parameters are pointers so that an explicit 0 (temperature 0 is the + // deterministic setting) can be told apart from a parameter the client did not send. + Temperature *float64 `json:"temperature,omitempty"` + TopP *float64 `json:"top_p,omitempty"` + TopK *int `json:"top_k,omitempty"` + StopSequences []string `json:"stop_sequences,omitempty"` + Stream bool `json:"stream,omitempty"` + System interface{} `json:"system,omitempty"` // Can be string or array of system messages + Thinking interface{} `json:"thinking,omitempty"` // Claude's thinking/extended thinking parameter + Tools []ClaudeTool `json:"tools,omitempty"` + ToolChoice interface{} `json:"tool_choice,omitempty"` } // ClaudeTool represents a tool definition in Claude format