feat: MCP server RAG/NL2SQL/PID 개선

Co-Authored-By: Claude Opus 4.8 <noreply@anthropic.com>
2026-06-10 08:11:44 +09:00
parent c850948630
commit 19c8c2e95c
4 changed files with 75 additions and 7 deletions
--- a/mcp-server/server.py
+++ b/mcp-server/server.py
@@ -61,6 +61,9 @@ KB_COLLECTIONS  = {

 # PostgreSQL 연결
 DB_CONNECTION_STRING = os.environ.get("DB_CONNECTION_STRING", "postgresql://postgres:postgres@localhost:5432/iiot_platform")
+# 라이브 데이터는 hc900 스키마에 있음. search_path 미지정 시 기본 public(레거시/stale)로 해석되므로 강제 지정.
+if "search_path" not in DB_CONNECTION_STRING and "options=" not in DB_CONNECTION_STRING:
+    DB_CONNECTION_STRING += ("&" if "?" in DB_CONNECTION_STRING else "?") + "options=-csearch_path%3Dhc900"
 DB_TIMEOUT = int(os.environ.get("DB_TIMEOUT", "10"))

 # C# McpClient(localhost:5001)와 통신: json_response+stateless로 단순 POST→JSON 방식
@@ -707,7 +710,8 @@ Views:
  v_tag_summary(base_tag TEXT, pv TEXT, sp TEXT, op TEXT, description TEXT, area TEXT, sub_area TEXT)

 Rules:
-  - SELECT only. tagname lowercase exact match.
+  - SELECT only. tagname UPPERCASE exact match (e.g. 'FICQ-6113.PV').
+  - If user input is lowercase, convert to UPPERCASE before querying.
  - value is TEXT; cast ::double precision when aggregating.
  - time_bucket() banned. For N-min buckets: to_timestamp(FLOOR(EXTRACT(EPOCH FROM recorded_at)/(N*60))*(N*60))
  - KST input = UTC-9 in DB.
@@ -1352,7 +1356,8 @@ async def query_with_nl(question: str) -> str:
        "You are a PostgreSQL SQL expert.\n"
        "Convert the user's question into a SELECT SQL.\n"
        "Return ONLY the SQL. No explanation, no markdown, NO <think> tags.\n"
-        "Use PostgreSQL syntax. tagname lowercase exact match.\n"
+        "Use PostgreSQL syntax. tagname UPPERCASE exact match (e.g. 'FICQ-6113.PV').\n"
+        "If user input is lowercase, convert to UPPERCASE before querying.\n"
        "value is TEXT; cast ::double precision when aggregating.\n"
        "KST input = UTC-9. Example: KST 12:00 = UTC 03:00.\n"
        "For N-min buckets: to_timestamp(FLOOR(EXTRACT(EPOCH FROM recorded_at)/(N*60))*(N*60)).\n"
@@ -1622,7 +1627,7 @@ async def trace_connections(start_tag: str, direction: str = "downstream", max_d
            # 각 노드에 실시간 상태(pv) 부착 — 병렬 펌프 중 '실제 가동 중'인 것을 식별.
            # 예: F-6101A/B 상류에 P-6102(R-RUN)·P-6201(L-STOP)이 병렬이면 현재 공급원은 P-6102.
            if path:
-                pv_tags = [p["tag_no"].lower() + ".pv" for p in path]
+                pv_tags = [p["tag_no"].upper() + ".PV" for p in path]
                cur.execute(
                    "SELECT tagname, livevalue FROM realtime_table WHERE tagname = ANY(%s)",
                    (pv_tags,),
@@ -1632,7 +1637,7 @@ async def trace_connections(start_tag: str, direction: str = "downstream", max_d
                    m = re.match(r'\{\s*\d+\s*\|\s*([^|]+?)\s*\|', lv or '')
                    pv_map[tn[:-3]] = (m.group(1).strip() if m else (lv or None))
                for p in path:
-                    p["live_state"] = pv_map.get(p["tag_no"].lower())
+                    p["live_state"] = pv_map.get(p["tag_no"])

            return json.dumps({
                "success": True,
--- a/mcp-server/worker/nl2sql_worker.py
+++ b/mcp-server/worker/nl2sql_worker.py
@@ -32,6 +32,9 @@ import httpx
 # ── 설정 ─────────────────────────────────────────────────────────────────────

 DB_CONNECTION_STRING = os.environ.get("DB_CONNECTION_STRING", "postgresql://postgres:postgres@localhost:5432/iiot_platform")
+# 라이브 데이터는 hc900 스키마에 있음. search_path 미지정 시 기본 public(레거시/stale)로 해석되므로 강제 지정.
+if "search_path" not in DB_CONNECTION_STRING and "options=" not in DB_CONNECTION_STRING:
+    DB_CONNECTION_STRING += ("&" if "?" in DB_CONNECTION_STRING else "?") + "options=-csearch_path%3Dhc900"
 DB_TIMEOUT = int(os.environ.get("DB_TIMEOUT", "10"))

 VLLM_BASE_URL = os.environ.get("VLLM_BASE_URL", "http://localhost:8000/v1")
@@ -101,6 +104,61 @@ def _llm_client():
 # DB 스키마 + SQL system 프롬프트 — worker/sql_prompt.py 로 단일화(production+eval 공유)
 from sql_prompt import DB_SCHEMA, SQL_SYSTEM_PROMPT  # noqa: E402,F401

+# ── 태그명 대소문자 정규화 ────────────────────────────────────────────────────
+
+# DB에 실제로 저장된 태그명 목록 (대문자)
+_tag_cache: dict | None = None
+_tag_cache_lock: object | None = None  # placeholder — sync path uses _get_db_connection
+
+
+def _load_tag_names() -> set[str]:
+    """realtime_table에서 현재 태그명 목록을 로드 (대문자)."""
+    global _tag_cache
+    try:
+        conn = _get_db_connection()
+        with conn.cursor() as cur:
+            cur.execute("SELECT tagname FROM hc900.realtime_table WHERE tagname IS NOT NULL")
+            _tag_cache = {r[0] for r in cur.fetchall()}
+        conn.close()
+        return _tag_cache
+    except Exception:
+        if _tag_cache is None:
+            _tag_cache = set()
+        return _tag_cache
+
+
+def _normalize_tag_case(sql: str) -> str:
+    """SQL 내의 태그명을 실제 DB 태그명으로 정규화 (대소문자 불일치 해결).
+
+    LLM이 소문자('ficq-6118.pv')로 생성한 태그명을
+    실제 DB에 저장된 대문자('FICQ-6118.PV')로 변환.
+    """
+    if not _tag_cache:
+        _load_tag_names()
+
+    if not _tag_cache:
+        return sql
+
+    # 태그명 패턴: 알파벳+숫자+하이픈+접미사(.pv/.sp/.op 등)
+    # 예: ficq-6118.pv, FICQ-6118.PV, xv-6105.pv
+    tag_pattern = re.compile(r"\b([a-zA-Z][a-zA-Z0-9]*-?[a-zA-Z0-9]*)(\.[a-zA-Z0-9_]+)?\b")
+
+    def _replace_tag(m: re.Match) -> str:
+        base = m.group(1)
+        suffix = m.group(2) or ""
+        candidate = base.upper() + suffix
+        if candidate in _tag_cache:
+            return candidate
+        # 접미사만 소문자일 수 있음 (예: FICQ-6118.pv → FICQ-6118.PV)
+        candidate2 = base.upper() + suffix.upper()
+        if candidate2 in _tag_cache:
+            return candidate2
+        return m.group(0)
+
+    normalized = tag_pattern.sub(_replace_tag, sql)
+    return normalized
+
+
 async def _generate_sql(natural_language: str) -> str:
    """자연어를 SQL로 변환."""
    client = _llm_client()
@@ -121,6 +179,8 @@ async def _generate_sql(natural_language: str) -> str:
    if sql.startswith("```"):
        lines = sql.splitlines()
        sql = "\n".join(lines[1:-1] if lines[-1].strip() == "```" else lines[1:]).strip()
+    # 태그명 대소문자 정규화
+    sql = _normalize_tag_case(sql)
    return sql

 # ── NL2SQL 도구 구현 ─────────────────────────────────────────────────────────
--- a/mcp-server/worker/pid_worker.py
+++ b/mcp-server/worker/pid_worker.py
@@ -33,6 +33,9 @@ VLLM_BASE_URL        = os.environ.get("VLLM_BASE_URL", "http://localhost:8000/v1
 from config import get_vllm_model
 VLLM_MODEL           = get_vllm_model()
 DB_CONNECTION_STRING = os.environ.get("DB_CONNECTION_STRING", "postgresql://postgres:postgres@localhost:5432/iiot_platform")
+# 라이브 데이터는 hc900 스키마에 있음. search_path 미지정 시 기본 public(레거시/stale)로 해석되므로 강제 지정.
+if "search_path" not in DB_CONNECTION_STRING and "options=" not in DB_CONNECTION_STRING:
+    DB_CONNECTION_STRING += ("&" if "?" in DB_CONNECTION_STRING else "?") + "options=-csearch_path%3Dhc900"
 DB_TIMEOUT           = int(os.environ.get("DB_TIMEOUT", "10"))

 _SERVER_DIR = os.path.dirname(os.path.dirname(os.path.abspath(__file__)))
--- a/mcp-server/worker/sql_prompt.py
+++ b/mcp-server/worker/sql_prompt.py
@@ -20,13 +20,13 @@ PostgreSQL 시계열 데이터베이스 스키마
 ※ DCS 태그: SELECT WHERE tag_dcs=TRUE, 현장 계기: WHERE tag_dcs=FALSE AND category='instrument'

 테이블: history_table  (시계열 이력)
-  tagname     TEXT         - 태그명 (모두 소문자, 예: 'ficq-6113.pv') — 대소문자 구분
+  tagname     TEXT         - 태그명 (대문자, 예: 'FICQ-6113.PV') — 대소문자 구분
  node_id     TEXT         - OPC UA 노드 ID
  value       TEXT         - 측정값, 수치 연산 시 ::double precision 캐스트 필요
  recorded_at TIMESTAMPTZ  - 기록 시각(UTC), 스냅샷 주기 약 60초

 테이블: realtime_table  (실시간 최신값)
-  tagname     TEXT         - 태그명 (모두 소문자)
+  tagname     TEXT         - 태그명 (대문자)
  node_id     TEXT         - OPC UA 노드 ID
  livevalue   TEXT         - 현재값
  timestamp   TIMESTAMPTZ  - 최종 갱신 시각
@@ -134,7 +134,7 @@ SQL_SYSTEM_PROMPT = (
    "      SELECT to_timestamp(...) AT TIME ZONE 'Asia/Seoul' AS bucket, AVG(...) AS avg_val\n"
    "      FROM ... GROUP BY to_timestamp(...), tagname ORDER BY to_timestamp(...), tagname\n"
    "- value column is TEXT; cast with ::double precision only when aggregating.\n"
-    "- All tagnames are lowercase (e.g. 'ficq-6113.pv'). Match exactly.\n"
+    "- All tagnames are UPPERCASE (e.g. 'FICQ-6113.PV'). Match exactly.\n"
    "- PostgreSQL LIKE: dot has no special meaning, no escaping needed.\n"
    "- Return ONLY the SQL statement. No explanation, no markdown.\n\n"
    f"{DB_SCHEMA}"