From 9701d98c55020f78575c348bb065bdc999d7a472 Mon Sep 17 00:00:00 2001 From: lethanhson9901 Date: Tue, 9 Dec 2025 11:08:35 +0700 Subject: [PATCH] feat: Add structured output support for OpenAI to Gemini conversion - Add support for response_format parameter in OpenAI requests - Map json_object type to Gemini's responseMimeType: application/json - Map json_schema type to Gemini's responseSchema with schema definition - Update buildGeminiGenerationConfig to handle response_format parameter This enables users to request structured JSON outputs using OpenAI's response_format API when calling Gemini models through the converter. --- src/converters/strategies/OpenAIConverter.js | 86 +++++++++++--------- 1 file changed, 49 insertions(+), 37 deletions(-) diff --git a/src/converters/strategies/OpenAIConverter.js b/src/converters/strategies/OpenAIConverter.js index 680852f..0b64b7d 100644 --- a/src/converters/strategies/OpenAIConverter.js +++ b/src/converters/strategies/OpenAIConverter.js @@ -190,12 +190,12 @@ export class OpenAIConverter extends BaseConverter { const mergedClaudeMessages = []; for (let i = 0; i < claudeMessages.length; i++) { const currentMessage = claudeMessages[i]; - + if (mergedClaudeMessages.length === 0) { mergedClaudeMessages.push(currentMessage); } else { const lastMessage = mergedClaudeMessages[mergedClaudeMessages.length - 1]; - + // 如果当前消息的 role 与上一条消息的 role 相同，则合并 content 数组 if (lastMessage.role === currentMessage.role) { lastMessage.content = lastMessage.content.concat(currentMessage.content); @@ -346,7 +346,7 @@ export class OpenAIConverter extends BaseConverter { // 处理 OpenAI chunk 对象 if (typeof openaiChunk === 'object' && !Array.isArray(openaiChunk)) { const choice = openaiChunk.choices?.[0]; - if (!choice){ + if (!choice) { return null; } @@ -400,7 +400,7 @@ export class OpenAIConverter extends BaseConverter { // } // }); // } - + // // 如果有 function.arguments，说明是参数增量 // if (toolCall.function?.arguments) { // events.push({ @@ -445,8 +445,8 @@ export class OpenAIConverter extends BaseConverter { if (finishReason) { // 映射 finish_reason const stopReason = finishReason === "stop" ? "end_turn" : - finishReason === "length" ? "max_tokens" : - "end_turn"; + finishReason === "length" ? "max_tokens" : + "end_turn"; events.push({ type: "content_block_stop", @@ -545,20 +545,20 @@ export class OpenAIConverter extends BaseConverter { toGeminiRequest(openaiRequest) { const messages = openaiRequest.messages || []; const { systemInstruction, nonSystemMessages } = extractSystemMessages(messages); - + const processedMessages = []; let lastMessage = null; - + for (const message of nonSystemMessages) { const geminiRole = message.role === 'assistant' ? 'model' : message.role; - + if (geminiRole === 'tool') { // Save previous model response with functionCall if (lastMessage) { processedMessages.push(lastMessage); lastMessage = null; } - + // Get function name from message.name or via tool_call_id let functionName = message.name; if (!functionName && message.tool_call_id) { @@ -574,11 +574,11 @@ export class OpenAIConverter extends BaseConverter { } } } - + // Build functionResponse according to Gemini API spec const parsedContent = safeParseJSON(message.content); const contentStr = typeof parsedContent === 'string' ? parsedContent : JSON.stringify(parsedContent); - + processedMessages.push({ role: 'user', parts: [{ @@ -594,9 +594,9 @@ export class OpenAIConverter extends BaseConverter { lastMessage = null; continue; } - + let processedContent = this.processOpenAIContentToGeminiParts(message.content); - + // Add tool_calls as functionCall to parts if (message.tool_calls && Array.isArray(message.tool_calls)) { for (const toolCall of message.tool_calls) { @@ -610,25 +610,25 @@ export class OpenAIConverter extends BaseConverter { } } } - + if (lastMessage && lastMessage.role === geminiRole && !message.tool_calls && Array.isArray(processedContent) && processedContent.every(p => p.text) && Array.isArray(lastMessage.parts) && lastMessage.parts.every(p => p.text)) { lastMessage.parts.push(...processedContent); continue; } - + if (lastMessage) processedMessages.push(lastMessage); lastMessage = { role: geminiRole, parts: processedContent }; } if (lastMessage) processedMessages.push(lastMessage); - + const geminiRequest = { contents: processedMessages.filter(item => item.parts && item.parts.length > 0) }; - + if (systemInstruction) geminiRequest.systemInstruction = systemInstruction; - + if (openaiRequest.tools?.length) { geminiRequest.tools = [{ functionDeclarations: openaiRequest.tools.map(t => { @@ -646,14 +646,14 @@ export class OpenAIConverter extends BaseConverter { delete geminiRequest.tools; } } - + if (openaiRequest.tool_choice) { geminiRequest.toolConfig = this.buildGeminiToolConfig(openaiRequest.tool_choice); } - + const config = this.buildGeminiGenerationConfig(openaiRequest, openaiRequest.model); if (Object.keys(config).length) geminiRequest.generationConfig = config; - + return geminiRequest; } @@ -663,20 +663,20 @@ export class OpenAIConverter extends BaseConverter { processOpenAIContentToGeminiParts(content) { if (!content) return []; if (typeof content === 'string') return [{ text: content }]; - + if (Array.isArray(content)) { const parts = []; - + for (const item of content) { if (!item) continue; - + if (item.type === 'text' && item.text) { parts.push({ text: item.text }); } else if (item.type === 'image_url' && item.image_url) { const imageUrl = typeof item.image_url === 'string' ? item.image_url : item.image_url.url; - + if (imageUrl.startsWith('data:')) { const [header, data] = imageUrl.split(','); const mimeType = header.match(/data:([^;]+)/)?.[1] || 'image/jpeg'; @@ -688,10 +688,10 @@ export class OpenAIConverter extends BaseConverter { } } } - + return parts; } - + return []; } @@ -711,13 +711,25 @@ export class OpenAIConverter extends BaseConverter { /** * 构建Gemini生成配置 */ - buildGeminiGenerationConfig({ temperature, max_tokens, top_p, stop, tools }, model) { + buildGeminiGenerationConfig({ temperature, max_tokens, top_p, stop, tools, response_format }, model) { const config = {}; config.temperature = checkAndAssignOrDefault(temperature, 1); config.maxOutputTokens = checkAndAssignOrDefault(max_tokens, 65535); config.topP = checkAndAssignOrDefault(top_p, 0.95); if (stop !== undefined) config.stopSequences = Array.isArray(stop) ? stop : [stop]; - + + // Handle response_format + if (response_format) { + if (response_format.type === 'json_object') { + config.responseMimeType = 'application/json'; + } else if (response_format.type === 'json_schema' && response_format.json_schema) { + config.responseMimeType = 'application/json'; + if (response_format.json_schema.schema) { + config.responseSchema = response_format.json_schema.schema; + } + } + } + // Gemini 2.5 and thinking models require responseModalities: ["TEXT"] // But this parameter cannot be added when using tools (causes 400 error) const hasTools = tools && Array.isArray(tools) && tools.length > 0; @@ -727,7 +739,7 @@ export class OpenAIConverter extends BaseConverter { } else if (hasTools && model && (model.includes('2.5') || model.includes('thinking') || model.includes('2.0-flash-thinking'))) { console.log(`[OpenAI->Gemini] Skipping responseModalities for model ${model} because tools are present`); } - + return config; } /** @@ -822,7 +834,7 @@ export class OpenAIConverter extends BaseConverter { name: toolCall.function.name || '', args: {} }; - + if (toolCall.function.arguments) { try { functionCall.args = typeof toolCall.function.arguments === 'string' @@ -833,7 +845,7 @@ export class OpenAIConverter extends BaseConverter { functionCall.args = { partial: toolCall.function.arguments }; } } - + parts.push({ functionCall }); } } @@ -894,7 +906,7 @@ export class OpenAIConverter extends BaseConverter { if (openaiRequest.messages && openaiRequest.messages.length > 0) { responsesRequest.messages = openaiRequest.messages.map(msg => ({ role: msg.role, - content: typeof msg.content === 'string' + content: typeof msg.content === 'string' ? [{ type: 'input_text', text: msg.content }] : msg.content })); @@ -1031,7 +1043,7 @@ export class OpenAIConverter extends BaseConverter { if (delta.tool_calls && delta.tool_calls.length > 0) { for (const toolCall of delta.tool_calls) { const outputIndex = toolCall.index || 0; - + // 如果有 function.name，说明是工具调用开始 if (toolCall.function && toolCall.function.name) { events.push({ @@ -1047,7 +1059,7 @@ export class OpenAIConverter extends BaseConverter { type: "response.output_item.added" }); } - + // 如果有 function.arguments，说明是参数增量 if (toolCall.function && toolCall.function.arguments) { events.push({ @@ -1080,7 +1092,7 @@ export class OpenAIConverter extends BaseConverter { generateOutputItemDone(responseId), generateResponseCompleted(responseId) ); - + // 如果有 usage 信息，更新最后一个事件 if (openaiChunk.usage && events.length > 0) { const lastEvent = events[events.length - 1];