Files
javis_bot/evals/test_personal_notes.py
javis-bot c51e5b2f68 fix(notes): say 'save' in the first 120 characters so the planner plans the note
The planner and router read only the first 120 characters of a tool's
description. personalNotes started by describing the notes, so for
'앞으로 나한테는 짧게 대답해줘' the planner planned a plain reply and the
answer promised to be brief from now on without saving anything.

Eval (qwen3.6-35b-a3b, game PC): from-now-on cases 0/6 -> 6/6, the
one-off and note-shapes-answer cases still pass (15/15).

Co-Authored-By: Claude Opus 5.5 <noreply@anthropic.com>
2026-11-18 01:50:16 +09:00

83 lines
3.6 KiB
Python

"""
Eval: per-person standing instructions (src/jarvis/tools/builtin/personal_notes.spec.md).
- A saved instruction reaches the answer: with the note "end every answer with the word
Roger." the answer to an unrelated question ends with "Roger" (and without the note it does not,
so the check measures the note, not the model's habits).
- Asking for something to apply from now on saves a note; a one-off question does not.
The notes backend is a fake that keeps what it is given; the tool runs for real.
Run: EVAL_JUDGE_MODEL=qwen2.5:7b python -m pytest evals/test_personal_notes.py -v
(v2 brain setup: EVAL_COMPACT_PROMPT=1 EVAL_INTENT_JUDGE=1)
"""
import re
import pytest
from conftest import requires_judge_llm
class FakeNotes:
def __init__(self):
self.acts = []
def act(self, action, **kw):
self.acts.append((action, kw))
if action == "add":
return {"ok": True, "item": {"id": "1", "text": kw["text"]}}
if action == "list":
return {"ok": True, "items": []}
return {"ok": True, "item": {"id": kw.get("note"), "text": ""}}
def _run(mock_config, eval_db, eval_dialogue_memory, query, notes=None):
from helpers import JUDGE_BASE_URL, JUDGE_MODEL
from jarvis.memory.history import HistoryContext
from jarvis.reply import turn_state
from jarvis.reply.engine import run_reply_engine
mock_config.ollama_base_url = JUDGE_BASE_URL
mock_config.ollama_chat_model = JUDGE_MODEL
backend = FakeNotes()
turn_state.reset()
turn_state.set_user("100")
turn_state.set_notes(backend)
try:
reply = run_reply_engine(db=eval_db, cfg=mock_config, tts=None, text=query,
dialogue_memory=eval_dialogue_memory, history=HistoryContext(),
legacy_memory=False, default_language="Korean", person_notes=notes)
finally:
turn_state.reset()
print(f"\n notes [{JUDGE_MODEL}] {query!r} notes={notes}: acts={backend.acts} reply={(reply or '')[:200]!r}")
return backend, reply or ""
def _ends_with_roger(reply: str) -> bool:
return re.search(r"roger\W*$", reply.strip(), re.IGNORECASE) is not None
@pytest.mark.eval
@requires_judge_llm
class TestPersonalNotes:
def test_a_saved_instruction_shapes_the_answer(self, mock_config, eval_db, eval_dialogue_memory):
_, reply = _run(mock_config, eval_db, eval_dialogue_memory, "고양이는 하루에 몇 시간 자?",
notes=["End every answer with the word Roger."])
assert _ends_with_roger(reply), reply
def test_without_the_note_the_answer_does_not_end_that_way(self, mock_config, eval_db, eval_dialogue_memory):
_, reply = _run(mock_config, eval_db, eval_dialogue_memory, "고양이는 하루에 몇 시간 자?")
assert not _ends_with_roger(reply), reply
# The second phrasing is from the field (2026-11-18): no note was saved but the answer said
# "from now on I will answer briefly".
@pytest.mark.parametrize("query", ["앞으로 나한테는 반말로 대답해", "자비스 앞으로 나한테는 짧게 대답해줘"])
def test_asking_for_it_from_now_on_saves_a_note(self, mock_config, eval_db, eval_dialogue_memory, query):
backend, _ = _run(mock_config, eval_db, eval_dialogue_memory, query)
adds = [kw for a, kw in backend.acts if a == "add"]
assert adds and adds[0]["text"], backend.acts
def test_a_one_off_question_saves_nothing(self, mock_config, eval_db, eval_dialogue_memory):
backend, _ = _run(mock_config, eval_db, eval_dialogue_memory, "고양이는 하루에 몇 시간 자?")
assert not any(a == "add" for a, _ in backend.acts), backend.acts