106 lines
4.1 KiB
Python
106 lines
4.1 KiB
Python
"""Ebene 0, Konsens: Vollständigkeits-Invariante (kein Kandidat verschwindet
|
|
still), explizite Ablehnung, uni-Beleg-Regel (1 Beleg reicht)."""
|
|
|
|
import re
|
|
|
|
import db
|
|
import fake_agents
|
|
import korpus
|
|
import llm
|
|
from conftest import run_anlegen, topic_anlegen
|
|
|
|
|
|
def _seed(art="uni", n_quellen=1):
|
|
topic = topic_anlegen("k-test", art=art)
|
|
quellen = [db.insert("quellen", topic=topic, art="datei", titel=f"q{i}.txt",
|
|
snapshot=f"q{i}.txt", hash=f"h{i}",
|
|
status="extrahiert", rolle="stoff")
|
|
for i in range(n_quellen)]
|
|
ctx = llm.Kontext(run_anlegen(topic), topic, "minimax")
|
|
ctx.ebene = "korpus"
|
|
return topic, quellen, ctx
|
|
|
|
|
|
def _kandidat(topic, punkt, quelle):
|
|
return db.insert("soll", topic=topic, punkt=punkt, status="kandidat",
|
|
belege=db.j([{"quelle": quelle, "zitat": f"Zitat {punkt}"}]))
|
|
|
|
|
|
def _ids(prompt):
|
|
return [int(i) for i in re.findall(r"^(\d+): ", prompt.split("KANDIDATEN:")[1],
|
|
re.MULTILINE)]
|
|
|
|
|
|
async def test_nachrunde_ordnet_vergessene_zu(monkeypatch):
|
|
topic, (q,), ctx = _seed()
|
|
a = _kandidat(topic, "Alpha", q)
|
|
b = _kandidat(topic, "Beta", q)
|
|
c = _kandidat(topic, "Gamma", q)
|
|
|
|
def judge(prompt):
|
|
if "NACHRUNDE" in prompt: # nur die fehlende id wird angeboten
|
|
assert _ids(prompt) == [c] and "Alpha" in prompt
|
|
return [{"punkt": "Alpha", "kandidaten": [c]}] # wörtlich → Merge
|
|
return [{"punkt": "Alpha", "kandidaten": [a]},
|
|
{"punkt": "Beta", "kandidaten": [b]}] # c stillschweigend vergessen
|
|
|
|
monkeypatch.setitem(fake_agents._HANDLER, "Korpus-Soll-Konsens", judge)
|
|
assert await korpus._konsens(ctx) == 2 # uni: 1 Quelle reicht als Beleg
|
|
alpha = db.one("SELECT * FROM soll WHERE topic=? AND status='bestaetigt'"
|
|
" AND punkt='Alpha'", (topic,))
|
|
assert len(db.uj(alpha["belege"])) == 2 # Gamma-Beleg in Alpha aufgegangen
|
|
for i in (a, b, c):
|
|
assert db.one("SELECT status FROM soll WHERE id=?", (i,))["status"] == "gefaltet"
|
|
assert korpus.messen(ctx) == []
|
|
|
|
|
|
async def test_ablehnung_ist_explizit(monkeypatch):
|
|
topic, (q,), ctx = _seed()
|
|
a = _kandidat(topic, "Alpha", q)
|
|
m = _kandidat(topic, "Einführung ins Thema", q)
|
|
|
|
def judge(prompt):
|
|
return [{"punkt": "Alpha", "kandidaten": [a]},
|
|
{"abgelehnt": True, "grund": "Meta-Überschrift", "kandidaten": [m]}]
|
|
|
|
monkeypatch.setitem(fake_agents._HANDLER, "Korpus-Soll-Konsens", judge)
|
|
assert await korpus._konsens(ctx) == 1
|
|
assert db.one("SELECT status FROM soll WHERE id=?", (m,))["status"] == "abgelehnt"
|
|
assert korpus.messen(ctx) == [] # abgelehnt ist erledigt, kein Befund
|
|
|
|
|
|
async def test_hartnaeckig_vergessene_wird_befund(monkeypatch):
|
|
topic, (q,), ctx = _seed()
|
|
a = _kandidat(topic, "Alpha", q)
|
|
c = _kandidat(topic, "Gamma", q)
|
|
|
|
def judge(prompt):
|
|
if "NACHRUNDE" in prompt:
|
|
return [] # Judge verweigert die Zuordnung dauerhaft
|
|
return [{"punkt": "Alpha", "kandidaten": [a]}]
|
|
|
|
monkeypatch.setitem(fake_agents._HANDLER, "Korpus-Soll-Konsens", judge)
|
|
await korpus._konsens(ctx)
|
|
assert db.one("SELECT status FROM soll WHERE id=?", (c,))["status"] == "kandidat"
|
|
befunde = korpus.messen(ctx)
|
|
assert [b["art"] for b in befunde] == ["soll_kandidat_offen"]
|
|
assert befunde[0]["item"] == str(c)
|
|
|
|
|
|
async def test_thema_braucht_zwei_quellen(monkeypatch):
|
|
topic, (q1, q2), ctx = _seed(art="thema", n_quellen=2)
|
|
x1 = _kandidat(topic, "X", q1)
|
|
x2 = _kandidat(topic, "X (Synonym)", q2)
|
|
y = _kandidat(topic, "Y", q1)
|
|
|
|
def judge(prompt):
|
|
if "NACHRUNDE" in prompt:
|
|
return []
|
|
return [{"punkt": "X", "kandidaten": [x1, x2]},
|
|
{"punkt": "Y", "kandidaten": [y]}]
|
|
|
|
monkeypatch.setitem(fake_agents._HANDLER, "Korpus-Soll-Konsens", judge)
|
|
assert await korpus._konsens(ctx) == 1 # X: 2 Quellen; Y: nur 1 → kein Punkt
|
|
assert db.one("SELECT status FROM soll WHERE id=?", (y,))["status"] == "kandidat"
|
|
assert {b["art"] for b in korpus.messen(ctx)} == {"soll_kandidat_offen"}
|