Initial commit: ABAP indexing pipeline (ingest, parser, summarize, index, query, wiki)

Co-Authored-By: Claude Fable 5.1 <noreply@anthropic.com>

tests/test_summarize_runner.py

@@ -0,0 +1,206 @@
+"""summarize.runner — 프로그램 요약 + unit 조각 추출의 상태·실패 사유·실행로그 라이프사이클.
+
+- 프로그램 요약이 먼저 생성되고 program.summary_json 에 저장된다
+- 실패 unit → summary_error 에 예외명+메시지, 재실행 시 실패분만 재시도
+- 조각은 logic_chunk / chunk_fts 에 저장되고 unit.chunk_count 에 반영된다
+- --fake 는 FakeLLM 으로 파이프라인 전체(조각 1개/unit)를 통과한다
+"""
+from __future__ import annotations
+
+import pytest
+from fastapi.testclient import TestClient
+
+from config.settings import settings
+from index.db import connect, loads
+import summarize.runner as runner
+
+PAYLOAD = {
+    "MAIN_PROGRAM": "ZRUNNER_T1",
+    "DESCRIPTION": "러너 테스트",
+    "TEXT_SYMBOL": [{"SYMBOL": "001", "TEXT": "회사코드 조회"}],
+    "INCLUDE_PROGRAM": [
+        {
+            "INCLUDE": "ZRUNNER_T1",
+            "SOURCE_CODE": (
+                "REPORT zrunner_t1.\n"
+                "START-OF-SELECTION.\n"
+                "  PERFORM ok_one.\n"
+                "  PERFORM fail_me.\n"
+                "FORM ok_one.\n"
+                "  SELECT * FROM t001 INTO TABLE gt_t001.\n"
+                "  SORT gt_t001 BY bukrs.\n"
+                "ENDFORM.\n"
+                "FORM fail_me.\n"
+                "  WRITE 2.\n"
+                "ENDFORM."
+            ),
+        }
+    ],
+}
+
+
+class StubLLM:
+    """FAIL_ME unit 에서만 예외. ok_one 에는 SELECT 조각 하나를 (줄 번호를 틀리게) 돌려준다."""
+
+    def __init__(self, fail_names: set[str]):
+        self.fail_names = fail_names
+        self.usage = {"calls": 0, "prompt_tokens": 0, "completion_tokens": 0, "cost_usd": 0.0}
+        self.prompts: list[str] = []
+
+    def complete_json(self, system: str, user: str) -> dict:
+        self.prompts.append(user)
+        for name in self.fail_names:
+            if f"name: {name}" in user:
+                raise RuntimeError("HTTP Error 429: rate limited")
+        self.usage["calls"] += 1
+        self.usage["prompt_tokens"] += 10
+        self.usage["completion_tokens"] += 20
+        if "[작업] program_summary" in user:
+            return {"program": "x", "business_purpose_ko": "회사코드 마스터 조회 테스트",
+                    "main_flow": ["회사코드 조회", "출력"], "business_tags": ["마스터관리"], "confidence": 0.7}
+        if "name: OK_ONE" in user:
+            return {"unit_purpose_ko": "회사코드 마스터를 읽는다",
+                    "chunks": [{"line_start": 7, "line_end": 8,   # 실제는 6~7 — 앵커로 보정되어야 함
+                                "first_line": "SELECT * FROM t001 INTO TABLE gt_t001.",
+                                "kind": "sql_select", "purpose_ko": "회사코드 마스터(T001) 전체 조회",
+                                "purpose_en": "Read company code master", "keywords_ko": ["회사코드"],
+                                "keywords_en": ["company code"], "sap_objects": ["T001"], "confidence": 0.9}]}
+        return {"unit_purpose_ko": "요약", "chunks": []}
+
+
+@pytest.fixture()
+def ingested(tmp_path, monkeypatch):
+    import query.api as api
+
+    monkeypatch.setattr(settings, "database_url", f"sqlite:///{tmp_path / 'index.db'}")
+    monkeypatch.setattr(settings, "data_normalized", tmp_path / "normalized")
+    monkeypatch.setattr(settings, "data_parsed", tmp_path / "parsed")
+    monkeypatch.setattr(settings, "llm_base_url", "")  # ingest 의 자동 요약은 끔
+    c = TestClient(api.app)
+    assert c.post("/ingest", json=PAYLOAD).json()["status"] == "loaded"
+    return c
+
+
+def _unit(con, name: str):
+    return con.execute(
+        "SELECT summary_status, summary_error, chunk_count, summary_json FROM unit "
+        "WHERE program='ZRUNNER_T1' AND name=?", (name,),
+    ).fetchone()
+
+
+def test_program_summary_chunks_failure_and_retry(ingested, monkeypatch):
+    stub = StubLLM({"FAIL_ME"})
+    monkeypatch.setattr(runner, "create_llm", lambda fake=False, **kw: stub)
+    stats = runner.summarize_units("ZRUNNER_T1", None, fake=False, dry_run=False, trigger="test")
+    assert stats["program_summaries"] == 1
+    assert stats["failed"] == 1
+    assert stats["done"] == 2          # OK_ONE + START-OF-SELECTION
+    assert stats["chunks"] == 1
+
+    # 프롬프트에 프로그램 요약 문맥과 텍스트 심볼·파서 힌트가 들어갔는가
+    unit_prompt = next(p for p in stub.prompts if "name: OK_ONE" in p)
+    assert "회사코드 마스터 조회 테스트" in unit_prompt      # 프로그램 요약 문맥
+    assert "L6 SELECT" in unit_prompt                         # 파서 힌트
+    assert "    6| " in unit_prompt                           # 줄번호 붙은 코드
+
+    con = connect()
+    try:
+        p = con.execute("SELECT summary_status, summary_json FROM program WHERE name='ZRUNNER_T1'").fetchone()
+        assert p["summary_status"] == "done"
+        assert loads(p["summary_json"])["prompt_version"] == runner.PROMPT_VERSION
+
+        fail = _unit(con, "FAIL_ME")
+        assert fail["summary_status"] == "failed"
+        assert "RuntimeError: HTTP Error 429" in fail["summary_error"]
+
+        ok = _unit(con, "OK_ONE")
+        assert ok["summary_status"] == "done" and ok["summary_error"] is None
+        assert ok["chunk_count"] == 1
+        thin = loads(ok["summary_json"])
+        assert thin["purpose_ko"] == "회사코드 마스터를 읽는다" and thin["chunk_count"] == 1
+
+        ch = con.execute("SELECT * FROM logic_chunk WHERE program='ZRUNNER_T1'").fetchone()
+        assert ch["chunk_id"].endswith("#FORM#OK_ONE#C1")
+        assert (ch["line_start"], ch["line_end"]) == (6, 7)   # 앵커로 보정됨
+        assert loads(ch["tables_read"]) == ["T001"]           # 파서가 채움
+        assert ch["kind"] == "sql_select"
+        # 검색 색인에 들어갔는가
+        hit = con.execute("SELECT chunk_id FROM chunk_fts WHERE chunk_fts MATCH '\"회사코드\"'").fetchone()
+        assert hit and hit["chunk_id"] == ch["chunk_id"]
+        # 얇은 unit 색인도 한 줄 요약으로 갱신
+        u = con.execute("SELECT purpose FROM unit_fts WHERE unit_id=?", (ch["unit_id"],)).fetchone()
+        assert u["purpose"] == "회사코드 마스터를 읽는다"
+
+        log = con.execute("SELECT * FROM llm_usage_log ORDER BY id DESC").fetchall()
+        assert len(log) == 1
+        assert log[0]["status"] == "done"
+        assert log[0]["total"] == stats["done"] + stats["failed"]
+        assert log[0]["failed"] == 1 and log[0]["chunks"] == 1
+        assert log[0]["trigger_by"] == "test"
+    finally:
+        con.close()
+
+    # 재실행(전부 성공) → 실패분만 재시도, 프로그램 요약은 스킵
+    stub2 = StubLLM(set())
+    monkeypatch.setattr(runner, "create_llm", lambda fake=False, **kw: stub2)
+    stats2 = runner.summarize_units("ZRUNNER_T1", None, fake=False, dry_run=False, trigger="test")
+    assert stats2["failed"] == 0 and stats2["done"] == 1 and stats2["program_summaries"] == 0
+    assert stats2["skipped"] == stats["done"]
+    assert not any("[작업] program_summary" in p for p in stub2.prompts)
+
+    con = connect()
+    try:
+        fail = _unit(con, "FAIL_ME")
+        assert fail["summary_status"] == "done" and fail["summary_error"] is None and fail["chunk_count"] == 0
+        assert con.execute("SELECT COUNT(*) FROM llm_usage_log").fetchone()[0] == 2
+    finally:
+        con.close()
+
+
+def test_all_skipped_writes_no_log(ingested, monkeypatch):
+    monkeypatch.setattr(runner, "create_llm", lambda fake=False, **kw: StubLLM(set()))
+    runner.summarize_units("ZRUNNER_T1", None, fake=False, dry_run=False, trigger="test")
+    stats = runner.summarize_units("ZRUNNER_T1", None, fake=False, dry_run=False, trigger="test")
+    assert stats["done"] == 0 and stats["failed"] == 0 and stats["program_summaries"] == 0
+    con = connect()
+    try:  # 두 번째 실행은 전부 스킵 — 로그 행이 추가되지 않는다
+        assert con.execute("SELECT COUNT(*) FROM llm_usage_log").fetchone()[0] == 1
+    finally:
+        con.close()
+
+
+def test_fake_llm_pipeline(ingested):
+    """--fake: FakeLLM 이 unit 마다 첫 실행문을 조각으로 만들어 파이프라인 전체를 통과한다."""
+    stats = runner.summarize_units("ZRUNNER_T1", None, fake=True, dry_run=False, trigger="test")
+    assert stats["failed"] == 0 and stats["program_summaries"] == 1
+    assert stats["chunks"] >= 1
+    con = connect()
+    try:
+        n = con.execute("SELECT COUNT(*) FROM logic_chunk WHERE program='ZRUNNER_T1'").fetchone()[0]
+        assert n == stats["chunks"]
+        for r in con.execute("SELECT c.*, u.line_start us, u.line_end ue FROM logic_chunk c "
+                             "JOIN unit u ON u.unit_id=c.unit_id"):
+            assert r["us"] <= r["line_start"] <= r["line_end"] <= r["ue"]
+    finally:
+        con.close()
+
+
+def test_reload_keeps_chunks_for_unchanged_units(ingested, monkeypatch):
+    """소스가 바뀐 unit 의 조각만 버리고, 같은 unit 의 조각은 재적재 후에도 남는다."""
+    monkeypatch.setattr(runner, "create_llm", lambda fake=False, **kw: StubLLM(set()))
+    runner.summarize_units("ZRUNNER_T1", None, fake=False, dry_run=False, trigger="test")
+
+    changed = dict(PAYLOAD)
+    changed["INCLUDE_PROGRAM"] = [dict(PAYLOAD["INCLUDE_PROGRAM"][0])]
+    changed["INCLUDE_PROGRAM"][0]["SOURCE_CODE"] = PAYLOAD["INCLUDE_PROGRAM"][0]["SOURCE_CODE"].replace(
+        "WRITE 2.", "WRITE 3.")   # FAIL_ME 만 바뀜
+    assert ingested.post("/ingest", json=changed).json()["status"] == "loaded"
+
+    con = connect()
+    try:
+        assert con.execute("SELECT COUNT(*) FROM logic_chunk WHERE unit_id LIKE '%#OK_ONE'").fetchone()[0] == 1
+        assert _unit(con, "OK_ONE")["summary_status"] == "done"
+        assert _unit(con, "FAIL_ME")["summary_status"] == "none"          # 재추출 대상
+        assert con.execute("SELECT summary_status FROM program WHERE name='ZRUNNER_T1'").fetchone()[0] == "stale"
+    finally:
+        con.close()