Jacob-9909 · Jacob-9909 · Sep 3, 2026 · Sep 2, 2026 · Sep 3, 2026
diff --git a/README.md b/README.md
@@ -144,7 +144,7 @@ uv run pytest
 * **Frontend**: Next.js 16 (React 19), TypeScript, Tailwind CSS v4
 * **Backend**: Python 3.12, FastAPI, Uvicorn, LangGraph, LangChain, Alembic
 * **Database**: PostgreSQL 17 (pgvector), Neo4j 5 Graph DB
-* **AI & Data**: NVIDIA NIM (`openai/gpt-oss-120b`, `BAAI/bge-m3`), Tavily, yfinance, 청약홈 공공데이터 API
+* **AI & Data**: NVIDIA NIM (`google/gemma-4-31b-it`, `BAAI/bge-m3`), Tavily, yfinance, 청약홈 공공데이터 API
 * **Infra**: 오라클 VM(systemd + Caddy/Let's Encrypt) · Vercel · GitHub Actions
 * **Tooling**: uv, ruff, pytest, ESLint
 

diff --git a/backend/app/api/tax_rates.py b/backend/app/api/tax_rates.py
@@ -11,6 +11,7 @@
 from pathlib import Path
 
 from fastapi import APIRouter, File, Form, HTTPException, UploadFile
+from pydantic import BaseModel
 
 from backend.app.api.uploads import read_upload_capped
 from backend.app.services.tax.rate_diff import diff_against_current
@@ -74,6 +75,22 @@ def _pdf_to_text(raw: bytes) -> str:
             ) from exc
 
 
+class ExtractTextRequest(BaseModel):
+    text: str
+    year: str = "2026"
+    use_llm: bool = True
+
+
+@router.post("/extract")
+def extract_text(req: ExtractTextRequest) -> dict:
+    """개정안 원문 텍스트에서 세율을 추출해 현행과 비교·검증한다."""
+    text = req.text.strip()
+    if not text:
+        raise HTTPException(status_code=400, detail="텍스트가 비어 있습니다.")
+    proposed = extract_rate_set(text[:_MAX_TEXT_CHARS], year=req.year, use_llm=req.use_llm)
+    return _diff_payload(proposed)
+
+
 @router.post("/extract/upload")
 async def extract_upload(
     file: UploadFile = File(...),

diff --git a/backend/app/services/agent/llm.py b/backend/app/services/agent/llm.py
@@ -182,7 +182,5 @@ def build_chat_model(temperature: float = 0.7, max_tokens: int = 4000) -> NIMCha
         max_tokens=max_tokens,
         timeout=REQUEST_TIMEOUT,
         max_retries=0,  # 같은 키로 재시도해봐야 429는 그대로 — 위 루프가 키를 바꿔가며 재시도한다
-        # 구 모델(qwen3-next-80b)은 스트리밍이 delta 4개로만 쪼개져 오면서 비스트리밍보다
-        # 2.5배 느려 disable_streaming=True로 껐었다. gpt-oss-120b로 교체 후 실측하니
-        # 정상 토큰 단위로 흐르므로(93청크/700토큰) 다시 켠다.
+        # 정상 토큰 단위로 스트리밍되는 gemma-4-31b-it 모델 사용
     )
diff --git a/backend/app/services/agent/nodes/intent.py b/backend/app/services/agent/nodes/intent.py
@@ -36,7 +36,10 @@
 
 tax_and_market_lookup을 골랐고 질문이 특정 자산 종류(예: 주식·채권·예금·부동산)에 한정되면,
 asset_types에 해당 자산 종류명을 적으십시오(세법 조회를 그 자산으로 좁힘). 자산을 특정하지 않은
-일반 질문이면 asset_types는 비워 두십시오(전체 세법 조회)."""
+일반 질문이면 asset_types는 비워 두십시오(전체 세법 조회).
+
+반드시 다음 JSON 형식(객체)으로만 응답하십시오:
+{"tools": ["도구명1", "도구명2"], "asset_types": [], "ticker": ""}"""
 
 
 # 데이터 검색이 필요함을 시사하는 토큰(짧은 메시지라도 이게 있으면 잡담이 아니다).
@@ -101,7 +104,14 @@ def _keyword_route(text: str) -> list[str]:
     if any(k in text for k in ("계산", "얼마야", "얼마나 내", "비과세")):
         route.append("tax_calculator")
     # 사기 메시지 검증
-    if any(k in text for k in ("사기", "피싱", "스미싱", "믿어도 되", "괜찮을까", "의심")):
+    if any(
+        k in text
+        for k in (
+            "사기", "피싱", "스미싱", "믿어도 되", "괜찮을까", "의심",
+            "사고났어", "이체해줘", "선입금", "수수료 먼저", "입금하면",
+            "보증금", "환급금", "계좌정보를 입력", "액정 깨져",
+        )
+    ):
         route.append("fraud_check")
     # 라이브 웹 리서치(현재 금리/상품·금리 동향·국세청 해석)
     if any(k in text for k in ("예금", "적금", "연금저축", "국채", "상품", "가입", "이율", "우대")):
@@ -181,7 +191,9 @@ class _Route(BaseModel):
         # 라우터 출력은 도구 이름 몇 개짜리 JSON이라 300토큰이면 충분하다. 기본값(4000)을
         # 두면 모델이 structured output에서 폭주할 때 4000토큰을 다 태우고서야 끝난다
         # (실측 45초). 상한을 낮추면 폭주해도 금방 끝나고 아래 _keyword_route로 폴백된다.
-        router = build_chat_model(temperature=0.0, max_tokens=300).with_structured_output(_Route)
+        router = build_chat_model(temperature=0.0, max_tokens=300).with_structured_output(
+            _Route, method="json_mode"
+        )
         prompt_input = f"[대화 맥락]\n{dialogue_context}\n\n[현재 사용자 발화]\n{user_text}"
         result = router.invoke(
             [SystemMessage(content=_INTENT_PROMPT), HumanMessage(content=prompt_input)]

diff --git a/docs/competition/proposal-draft.md b/docs/competition/proposal-draft.md
@@ -160,7 +160,7 @@
 
 - **Frontend**: Next.js 16(App Router), TypeScript, Tailwind, GSAP/D3/WebGL 시각화
 - **Backend**: FastAPI, LangGraph(StateGraph + PostgresSaver 체크포인터)
-- **AI**: NVIDIA NIM(gpt-oss-120b / bge-m3 임베딩), Tavily·Naver 웹검색 도구
+- **AI**: NVIDIA NIM(google/gemma-4-31b-it / bge-m3 임베딩), Tavily·Naver 웹검색 도구
 - **Data**: PostgreSQL(pgvector) · Neo4j · yfinance · 청약홈 공공데이터
 - **Infra**: Docker Compose, Alembic 마이그레이션, GitHub Actions CI
 - 전체 다이어그램: `architecture.svg` 참조

diff --git a/frontend/src/app/tax-rates/page.tsx b/frontend/src/app/tax-rates/page.tsx
@@ -1,14 +1,20 @@
 "use client";
 
 import { useRef, useState } from "react";
-import { ArrowRight, CheckCircle, Warning, FilePdf, ShieldCheck } from "@phosphor-icons/react";
+import { ArrowRight, CheckCircle, Warning, FilePdf, ShieldCheck, Sparkle } from "@phosphor-icons/react";
 import {
+  extractRates,
   extractRatesUpload,
   type RateDiffRow,
   type RateExtractResult,
 } from "@/lib/api";
 import { Card, PageTitle, SectionLabel, fmtKRW } from "@/components/ui";
 
+const SAMPLE_AMENDMENT = `2026년 귀속 세법개정안 요약
+- 해외주식 양도소득세율을 20%에서 22%로 상향한다.
+- 이자·배당 분리과세율은 15.4%로 유지한다.
+- 금융소득종합과세 기준을 2,000만원에서 3,000만원으로 상향한다.`;
+
 /** 세율값을 종류에 맞게 렌더 — rate는 퍼센트, amount는 원화. */
 function fmtRateValue(kind: "rate" | "amount", value: number): string {
   return kind === "rate" ? `${(value * 100).toFixed(2).replace(/\.?0+$/, "")}%` : fmtKRW(value);
@@ -33,12 +39,27 @@ function DiffRow({ row }: { row: RateDiffRow }) {
 }
 
 export default function TaxRatesPage() {
+  const [text, setText] = useState(SAMPLE_AMENDMENT);
   const [year, setYear] = useState("2026");
   const [result, setResult] = useState<RateExtractResult | null>(null);
   const [busy, setBusy] = useState(false);
   const [error, setError] = useState<string | null>(null);
   const fileRef = useRef<HTMLInputElement>(null);
 
+  async function onExtract() {
+    if (!text.trim()) return;
+    setBusy(true);
+    setError(null);
+    setResult(null);
+    try {
+      setResult(await extractRates(text, year, false));
+    } catch (e) {
+      setError(e instanceof Error ? e.message : "세율 추출에 실패했습니다.");
+    } finally {
+      setBusy(false);
+    }
+  }
+
   async function onUpload(file: File) {
     setBusy(true);
     setError(null);
@@ -58,12 +79,19 @@ export default function TaxRatesPage() {
       <PageTitle
         eyebrow="결정론 계산기 · 개정 대응"
         title="세율 개정안 인입 (미리보기)"
-        subtitle="개정안 PDF·파일을 올리면 시스템이 세율을 추출해 현행과 비교·검증합니다. 자동 반영은 하지 않습니다 — 세율은 코드 상수로만 결정되는 결정론 불변식을 지키기 위해서입니다. LLM은 추출만, 계산은 코드가 합니다."
+        subtitle="개정안 텍스트를 붙여넣거나 PDF·파일을 올리면 시스템이 세율을 추출해 현행과 비교·검증합니다. 자동 반영은 하지 않습니다 — 세율은 코드 상수로만 결정되는 결정론 불변식을 지키기 위해서입니다. LLM은 추출만, 계산은 코드가 합니다."
       />
 
-      {/* ── 입력 (파일 업로드) ── */}
+      {/* ── 입력 (텍스트 붙여넣기 + 파일 업로드) ── */}
       <Card className="space-y-4">
-        <SectionLabel>개정안 파일 업로드</SectionLabel>
+        <SectionLabel>개정안 텍스트 입력</SectionLabel>
+        <textarea
+          value={text}
+          onChange={(e) => setText(e.target.value)}
+          rows={5}
+          className="field w-full resize-y p-3 font-mono-spec text-sm"
+          placeholder="세법개정안 텍스트를 붙여넣으세요…"
+        />
         <div className="flex flex-wrap items-center gap-3">
           <label className="flex items-center gap-2 text-sm text-muted">
             귀속연도
@@ -74,6 +102,17 @@ export default function TaxRatesPage() {
             />
           </label>
 
+          <button
+            onClick={onExtract}
+            disabled={busy || !text.trim()}
+            className="btn-accent inline-flex items-center gap-2 disabled:opacity-50"
+          >
+            <Sparkle size={16} weight="fill" />
+            {busy ? "추출 중…" : "세율 추출"}
+          </button>
+
+          <span className="text-xs text-muted">또는</span>
+
           <input
             ref={fileRef}
             type="file"
@@ -87,27 +126,15 @@ export default function TaxRatesPage() {
           <button
             onClick={() => fileRef.current?.click()}
             disabled={busy}
-            className="btn-accent inline-flex items-center gap-2 disabled:opacity-50"
+            className="btn-ghost inline-flex items-center gap-2 disabled:opacity-50"
           >
             <FilePdf size={16} />
-            {busy ? "추출 중…" : "개정안 파일 업로드"}
+            개정안 파일 업로드
           </button>
         </div>
         <p className="text-xs text-muted">
-          PDF·TXT·MD 지원. PDF는 서버가 표까지 텍스트로 추출해 세율을 뽑고 현행과 비교합니다.
-          별도 텍스트 작성 없이 개정안 문서 그대로 올리면 됩니다.
-        </p>
-        <p className="text-xs text-muted">
-          세법개정안 원문은{" "}
-          <a
-            href="https://www.moef.go.kr"
-            target="_blank"
-            rel="noopener noreferrer"
-            className="text-accent underline decoration-accent/40 underline-offset-2 hover:decoration-accent"
-          >
-            기획재정부(moef.go.kr) ↗
-          </a>
-          에서 받아 그대로 올리세요.
+          위 기본 예시로 1초 만에 테스트하거나, PDF·TXT·MD 파일을 직접 올릴 수 있습니다.
+          PDF는 서버가 표까지 텍스트로 추출해 세율을 뽑고 현행과 비교합니다.
         </p>
       </Card>
 

diff --git a/frontend/src/lib/api.ts b/frontend/src/lib/api.ts
@@ -684,6 +684,14 @@ export function getCurrentRates(year = "2026"): Promise<RateCurrent> {
   return apiGet(`/api/v1/tax-rates/current?year=${encodeURIComponent(year)}`);
 }
 
+export function extractRates(
+  text: string,
+  year = "2026",
+  useLlm = false,
+): Promise<RateExtractResult> {
+  return apiPost("/api/v1/tax-rates/extract", { text, year, use_llm: useLlm });
+}
+
 /** 개정안 파일(PDF·TXT·MD)을 업로드해 추출한다. PDF는 서버가 파서로 텍스트를 뽑아 동일 파이프라인을 탄다. */
 export async function extractRatesUpload(
   file: File,

diff --git a/openwiki/.last-update.json b/openwiki/.last-update.json
@@ -1,6 +1,6 @@
 {
-  "updatedAt": "2026-08-30T14:24:28.855Z",
+  "updatedAt": "2026-09-02T14:22:01.117Z",
   "command": "update",
-  "gitHead": "3916b2f5a44de605bf2e604c7cd04cd7c459a2fd",
+  "gitHead": "8f27862e3e7e1d239b8010b65fea71e06f913907",
   "model": "openai/gpt-oss-120b"
 }
diff --git a/openwiki/api.md b/openwiki/api.md
@@ -6,6 +6,7 @@ This repository exposes a **FastAPI** server under the `/api/v1` prefix. The API
 
 * **Path Prefix** – All routes start with `/api/v1`.
 * **Response Models** – Pydantic models are used for request validation and response schemas (see the source files).
+* **Rate Limiting** – Global IP‑based request throttling is enforced by `backend/app/middleware/rate_limit.py`. See the middleware source for configuration (`RATE_LIMIT`, `RATE_LIMIT_WINDOW`).
 * **Authentication** – Auth routes are available (`/api/v1/auth/login`). The UI can obtain a JWT token via login; if `AUTH_ENABLED` is false, the auth layer is bypassed and `user_uuid` may be passed directly.
 * **Streaming** – The chat endpoint supports Server‑Sent Events (SSE) for token‑by‑token streaming.
 
@@ -53,18 +54,6 @@ This repository exposes a **FastAPI** server under the `/api/v1` prefix. The API
 ---
 
 ## Cheongyak (Housing) Endpoints (`backend/app/api/cheongyak.py`)
-
-## Tax Rates API (`backend/app/api/tax_rates.py`)
-
-| Method | Path | Description |
-|--------|------|-------------|
-| `POST` | `/tax-rates/extract` | Extract tax rate proposals from raw text (up to 20k chars). Returns diff against current rates and validation issues. |
-| `POST` | `/tax-rates/extract/upload` | Upload a `.txt`, `.md`, or `.pdf` file to extract tax rates. Supports same diff and validation as above. |
-| `POST` | `/tax-rates/apply` | Apply a validated tax rate proposal to the overlay for a given year. Returns confirmation and active rates payload. |
-| `GET` | `/tax-rates/current` | Retrieve the current effective tax rates for a given year (default 2026), after any overlays. |
-| `GET` | `/tax-rates/current?year=2025` | Retrieve rates for a specific year.
-
-These endpoints support the tax calculator tool and UI tax‑rate management features. Source: `backend/app/api/tax_rates.py`.
 | Method | Path | Description |
 |--------|------|-------------|
 | `GET` | `/cheongyak/list/{kind}` | List housing projects of a given `kind` (e.g., `apt`, `officetel`).
@@ -75,28 +64,29 @@ These endpoints support the tax calculator tool and UI tax‑rate management fea
 
 ---
 
-## Finetuning & Embedding Pipelines (`backend/app/api/finetune.py`)
+## Tax Rates API (`backend/app/api/tax_rates.py`)
 | Method | Path | Description |
 |--------|------|-------------|
-| `POST` | `/finetune/upload` | Upload a document (PDF/TXT/MD/JSONL) for embedding and triplet generation.
-| `POST` | `/finetune/jobs` | Start a new finetuning pipeline job. Returns a job ID.
-| `POST` | `/finetune/train/jobs` | Start LoRA finetuning training job on generated triplet dataset. Returns a job ID.
-| `GET` | `/finetune/jobs` | List all finetuning jobs with status.
-| `GET` | `/finetune/jobs/{id}` | Get detailed progress and logs for a specific job.
-| `GET` | `/finetune/datasets` | Preview generated training/evaluation datasets (triplets) for a given sub‑directory.
+| `POST` | `/tax-rates/extract` | Extract tax rate proposals from raw text (up to 20k chars). Returns diff against current rates and validation issues. |
+| `POST` | `/tax-rates/extract/upload` | Upload a `.txt`, `.md`, or `.pdf` file to extract tax rates. Supports same diff and validation as above. |
+| `POST` | `/tax-rates/apply` | Apply a validated tax rate proposal to the overlay for a given year. Returns confirmation and active rates payload. |
+| `GET` | `/tax-rates/current` | Retrieve the current effective tax rates for a given year (default 2026), after any overlays. |
+| `GET` | `/tax-rates/current?year=2025` | Retrieve rates for a specific year. |
 
 ---
 
-## Knowledge Graph Endpoints (`backend/app/api/graph.py`)
+## Knowledge Graph & Document RAG Endpoints (`backend/app/api/graph.py`)
 | Method | Path | Description |
 |--------|------|-------------|
-| `POST` | `/graph/build/jobs` | Trigger an incremental Neo4j graph build job.
-| `GET` | `/graph/build/jobs` | List active/completed graph‑build jobs.
-| `GET` | `/graph/build/jobs/{id}` | Retrieve job status and logs.
-| `GET` | `/graph/snapshot` | Return a JSON snapshot of Neo4j nodes/edges (used by the UI visualiser).
-| `GET` | `/graph/documents` | List currently ingested RAG documents (used by Knowledge Panel).
-| `POST` | `/graph/ingest/jobs` | Trigger a document ingest job (parse & embed uploaded file).
-| `GET` | `/graph/ingest/jobs/{job_id}` | Get status of a document ingest job.
+| `GET` | `/graph/documents` | List currently ingested RAG documents (used by Knowledge Panel). |
+| `DELETE` | `/graph/documents/{source}` | Delete a document from RAG (`emb_passages` records and `data/raw_documents/` file). |
+| `POST` | `/graph/upload` | Upload a financial document to `data/raw_documents/` for embedding. |
+| `POST` | `/graph/ingest/jobs` | Trigger a document ingest job (parse & embed uploaded file). |
+| `GET` | `/graph/ingest/jobs/{job_id}` | Get status of a document ingest job. |
+| `POST` | `/graph/build/jobs` | Trigger an incremental Neo4j graph build job. |
+| `GET` | `/graph/build/jobs` | List active/completed graph‑build jobs. |
+| `GET` | `/graph/build/jobs/{id}` | Retrieve job status and logs. |
+| `GET` | `/graph/snapshot` | Return a JSON snapshot of Neo4j nodes/edges (used by the UI visualiser). |
 
 
 ---

diff --git a/openwiki/architecture.md b/openwiki/architecture.md
@@ -11,7 +11,7 @@ Midas Touch is organized around **four major technical domains** that align with
 
 ## Backend Architecture
 
-* **Entry point** – `backend/app/main.py` creates a FastAPI app, registers eight routers (`chat`, `users`, `finetune`, `graph`, `query`, `stocks`, `cheongyak`, `research`). It also starts a background validation loop that periodically scores the analysis memory.
+* **Entry point** – `backend/app/main.py` creates a FastAPI app, registers routers (`auth`, `chat`, `users`, `tax_rates`, `graph`, `query`, `stocks`, `cheongyak`, `research`). It also starts a background validation loop that periodically scores the analysis memory.
 * **API Routers** – each router lives under `backend/app/api/` and groups related endpoints (e.g., `stocks.py` exposes quick‑analysis, backtest, grid‑search, and memory statistics). Routes are mounted under `/api/v1`.
 * **Services** – business logic is encapsulated in `backend/app/services/`:
   * `agent/` – LangGraph state graph, node definitions, tool implementations, persistence via `PostgresSaver`.