Add Langfuse as a self-hosted eval and trace viewer

2026-08-21 10:35:25 +07:00
parent 4490a1abf0
commit 53582b6030
12 changed files with 1129 additions and 280 deletions
@@ -0,0 +1,92 @@
+import { NextResponse } from "next/server";
+import type { ChatMessage } from "@duoc-thu/shared-types";
+import { buildAssistantMessage, type RagAnswerRaw } from "../_lib/ragResponse";
+
+export const runtime = "nodejs";
+
+// Same resolution rule as `/api/chat` and `/api/history` — RAG never goes
+// through api-gateway (it only proxies `/auth/*`), so this always talks to
+// ai-service directly.
+const AI_SERVICE_URL = process.env.AI_SERVICE_URL ?? "http://localhost:8000";
+
+interface TranscriptMessageRaw {
+  role: "user" | "assistant";
+  trace_id: string;
+  content: string | null;
+  decision?: string | null;
+  reason?: string | null;
+  resolved_drug_id?: string | null;
+  citations?: RagAnswerRaw["citations"];
+  generated?: boolean;
+  quick_replies?: string[];
+  blocks?: RagAnswerRaw["blocks"];
+  answer_mode?: RagAnswerRaw["answer_mode"];
+  answer_plan?: RagAnswerRaw["answer_plan"];
+  candidate_assessments?: RagAnswerRaw["candidate_assessments"];
+  disclaimer?: string | null;
+  created_at: string;
+}
+
+interface TranscriptResponseRaw {
+  messages: TranscriptMessageRaw[];
+}
+
+function toChatMessage(row: TranscriptMessageRaw): ChatMessage {
+  if (row.role === "user") {
+    return {
+      id: `user-${row.trace_id}`,
+      role: "user",
+      content: row.content ?? "",
+      createdAt: row.created_at,
+    };
+  }
+  // Reuses the exact live-turn mapping (`/api/chat`'s `buildAssistantMessage`)
+  // so a replayed answer renders identically to the one the user originally
+  // saw — same abstain/refusal fallback, same citation grouping.
+  const raw: RagAnswerRaw = {
+    decision: row.decision ?? "answerable",
+    reason: row.reason ?? "",
+    answer: row.content,
+    resolved_drug_id: row.resolved_drug_id ?? null,
+    citations: row.citations ?? [],
+    generated: row.generated,
+    quick_replies: row.quick_replies,
+    blocks: row.blocks,
+    answer_mode: row.answer_mode,
+    answer_plan: row.answer_plan,
+    candidate_assessments: row.candidate_assessments,
+    disclaimer: row.disclaimer,
+  };
+  return buildAssistantMessage(raw, row.trace_id, row.created_at);
+}
+
+export async function GET(request: Request) {
+  const { searchParams } = new URL(request.url);
+  const conversationId = searchParams.get("conversation_id")?.trim() || "";
+
+  if (!conversationId) {
+    return NextResponse.json({ messages: [] });
+  }
+
+  try {
+    const base = AI_SERVICE_URL.replace(/\/v1\/rag(?:\/query)?\/?$/, "");
+    const targetUrl = `${base}/v1/rag/transcript?conversation_id=${encodeURIComponent(conversationId)}`;
+
+    const upstream = await fetch(targetUrl, {
+      method: "GET",
+      headers: { "X-Client-Version": "1.0.0" },
+      cache: "no-store",
+      signal: AbortSignal.timeout(8_000),
+    });
+
+    if (!upstream.ok) {
+      return NextResponse.json({ messages: [] });
+    }
+
+    const data = (await upstream.json()) as TranscriptResponseRaw;
+    const messages = (data.messages ?? []).map(toChatMessage);
+    return NextResponse.json({ messages });
+  } catch {
+    return NextResponse.json({ messages: [] });
+  }
+}