Research Agent MCP

import { ANTHROPIC_SEARCH_SUGGESTIONS_PROMPT } from "@/lib/system-prompt" import Anthropic from "@anthropic-ai/sdk" import type { ToolUnion } from "@anthropic-ai/sdk/resources/index.mjs" import type { NextRequest } from "next/server" import type { AnthropicModelMessage, ModelConfig } from "@/lib/types" import { DEFAULT_MODEL_CONFIG } from "@/lib/models" export const runtime = "nodejs" const anthropic = new Anthropic({ apiKey: process.env.ANTHROPIC_API_KEY || "", }) function extractSearchSuggestionsFromText( text: string, ): { searchTerms: string[]; confidence: number; reasoning: string } | null { try { const regex = /```SEARCH_TERMS_JSON\s*({[\s\S]*?})\s*```/ const match = text.match(regex) if (match && match[1]) { const jsonStr = match[1].trim() const searchSuggestions = JSON.parse(jsonStr) if ( searchSuggestions && Array.isArray(searchSuggestions.searchTerms) && typeof searchSuggestions.confidence === "number" && typeof searchSuggestions.reasoning === "string" ) { return { searchTerms: searchSuggestions.searchTerms, confidence: searchSuggestions.confidence, reasoning: searchSuggestions.reasoning, } } } return null } catch (error) { console.error("Error extracting search suggestions:", error) return null } } function removeSearchTermsJson(text: string): string { const pattern = /```SEARCH_TERMS_JSON\s*({[\s\S]*?})\s*```/g return text.replace(pattern, "").trim() } function validateMessages(messages: any[]): AnthropicModelMessage[] { return messages .filter((message) => { const isValid = message && typeof message === "object" && (message.role === "user" || message.role === "assistant") && // Only user or assistant roles typeof message.content === "string" if (!isValid) { console.warn("Invalid message format detected and filtered out:", message) } return isValid }) .map((message) => ({ id: message.id || nanoid(), role: message.role as "user" | "assistant", content: message.content, })) } // Simple nanoid implementation for generating IDs function nanoid(size = 21): string { const urlAlphabet = "ModuleSymbhasOwnPr-0123456789ABCDEFGHNRVfgctiUvz_KqYTJkLxpZXIjQW" let id = "" let i = size while (i--) { id += urlAlphabet[(Math.random() * 64) | 0] } return id } export async function POST(req: NextRequest) { try { const body = await req.json() if (!body || !Array.isArray(body.messages)) { return new Response(JSON.stringify({ error: "Invalid request body. Expected messages array." }), { status: 400, headers: { "Content-Type": "application/json" }, }) } let selectedModel = "claude-3-5-sonnet-latest" if (body.model && typeof body.model === "string") { selectedModel = body.model } else { const modelCookie = req.cookies.get("selectedAnthropicModel") if (modelCookie) { selectedModel = modelCookie.value } } const modelConfig: ModelConfig = body.modelConfig || DEFAULT_MODEL_CONFIG const reasoningType = body.reasoningType || "Reasoning" const validatedMessages = validateMessages(body.messages) const formattedMessages = validatedMessages.map((msg) => ({ role: msg.role, content: msg.content, })) const tools = [ { name: "web_search", type: "web_search_20250305", max_uses: 1, cache_control: { type: "ephemeral" } }, ] as ToolUnion[] const encoder = new TextEncoder() const readableStream = new ReadableStream({ async start(controller) { try { const stream = await anthropic.messages.stream({ model: selectedModel, max_tokens: modelConfig.maxTokens, temperature: modelConfig.temperature, //top_p: modelConfig.topP, system: ANTHROPIC_SEARCH_SUGGESTIONS_PROMPT, messages: formattedMessages, tools, // claude sonnet 3.7 에서 다음을 지원 // thinking: { type: 'enabled', budget_tokens: 1600 }, // thinking 을 사용하려면 temperature 가 1로 설정되어야 하고 // top P 는 파라미터에서 사용되어서는 안된다. // claude sonnet 3.5는 thinking 을 지원하지 않으니 thinking 이 설정되어서는 안된다. // thinking: { type: 'enabled', budget_tokens: 1600 }, }) const sources: any[] = [] let fullText = "" console.log("Starting Anthropic stream execution with model:", selectedModel, "reasoningType:", reasoningType) for await (const chunk of stream) { if (chunk.type === "content_block_delta") { if (chunk.delta.type === "text_delta") { fullText += chunk.delta.text const data = { type: "text-delta", text: chunk.delta.text, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } else if (chunk.delta.type === "citations_delta") { const citation = chunk.delta.citation if (citation.type === "web_search_result_location") { const source = { url: citation.url || "", title: citation.title || "", cited_text: citation.cited_text || "", } const existingSourceIndex = sources.findIndex((s) => s.url === source.url) if (existingSourceIndex >= 0) { if (source.cited_text && !sources[existingSourceIndex].cited_text.includes(source.cited_text)) { sources[existingSourceIndex].cited_text += `\n\n${source.cited_text}` } } else { sources.push(source) } } } else if (chunk.delta.type === "thinking_delta") { console.log(`Claude Thinking: ${chunk.delta.thinking}`) } } else if (chunk.type === "message_delta") { if (chunk.delta.stop_reason) { const data = { type: "stop_reason", stop_reason: chunk.delta.stop_reason, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } if (chunk.usage) { const data = { type: "usage", usage: chunk.usage, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } } } if (sources.length > 0) { const data = { type: "sources", sources, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } const searchSuggestions = extractSearchSuggestionsFromText(fullText) if (searchSuggestions) { const data = { type: "searchSuggestions", searchSuggestions: searchSuggestions.searchTerms, confidence: searchSuggestions.confidence, reasoning: searchSuggestions.reasoning, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } const cleanedText = removeSearchTermsJson(fullText) if (cleanedText !== fullText) { const data = { type: "cleaned-text", text: cleanedText, messageId: Date.now().toString(), } controller.enqueue(encoder.encode(`data: ${JSON.stringify(data)}\n\n`)) } const configData = { type: "model-config", config: modelConfig, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(configData)}\n\n`)) const selectedModelData = { type:"selected-model", model: selectedModel, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(selectedModelData)}\n\n`)) const reasoningTypeData = { type:"reasoning-type", reasoning: reasoningType, } controller.enqueue(encoder.encode(`data: ${JSON.stringify(reasoningTypeData)}\n\n`)) controller.close() } catch (error) { console.error("Error in stream processing:", error) const errorData = { type: "error", error: error instanceof Error ? error.message : String(error), } controller.enqueue(encoder.encode(`data: ${JSON.stringify(errorData)}\n\n`)) controller.close() } }, }) return new Response(readableStream, { headers: { "Content-Type": "text/event-stream", "Cache-Control": "no-cache", Connection: "keep-alive", }, }) } catch (error) { console.error("Error in Anthropic chat API:", error) return new Response(JSON.stringify({ error: "Failed to process Anthropic chat request" }), { status: 500, headers: { "Content-Type": "application/json" }, }) } }