7cd2fa5cf4
Measured first, by the ped-ai session, fifteen runs of five prompts with the gateway cache bypassed. Retrieval was already deterministic: identical shortlist and identical scores every time, and the citation checker stripped none of the 45 markers written — invented citations are not the problem here. Generation was the whole variance. At temperature 0.3 the same sources and the same prompt gave answers differing by 15-70% of their text; one differential swung between a 35-word uncited paraphrase and a 180-word cited list. So temperature 0 and a seed. Temperature 0 alone was not enough — three runs still differed — and temperature 0 with a fixed seed came back byte-identical. The seed is derived from the question, normalised for case and spacing, so two people asking the same thing get the same answer and a different question is not pinned to the same sample. An empty reply is asked once more before it becomes a 502. One in fifteen came back empty from a healthy model in 4.9 seconds — not a refusal, not an error, just nothing. A short query that finds almost nothing is retried against the nearest article title. "kawasaki criteria" finds fourteen sources; "kawasaki critera" found none — the lexical ranker cannot match a token that is in no index, and the embedding of a misspelling is not near the embedding of the word. Trigrams do not care: that typo scores 0.36 against "Kawasaki disease" with the next article at 0.11, and the gap is what makes it safe to act on. pg_trgm is created at startup beside vector, with a migration for the record. And an answer drawn from the library must cite it. Not a hallucination guard — nothing was stripped in fifteen runs — but one answer used the sources and cited none of them, which leaves the learner an assertion and nowhere to check it. Also, article drafts are weighted towards mechanism, in the wording the ped-ai rewriter is using, so the two lanes read alike: why the body does what it does, with features and management explained through it rather than listed. Figure lines and cross-references survive a refine. Co-Authored-By: Claude Opus 5 <noreply@anthropic.com> Claude-Session: https://claude.ai/code/session_01TqXevQJhxFrM7jJg82cgZN
361 lines
19 KiB
Python
361 lines
19 KiB
Python
"""AI Mode: retrieval, the citation contract, and thread ownership.
|
|
|
|
Disposable SQLite; the model itself is stubbed, because what is worth testing
|
|
here is not what a model says but what the server does with it. The safety
|
|
property — an invented citation cannot survive — has to hold whatever comes back.
|
|
"""
|
|
import unittest
|
|
from unittest.mock import AsyncMock, patch
|
|
|
|
import test_quiz_builder as fixtures
|
|
from app.models.article import Article, ArticleSectionIndex, QuestionArticleLink
|
|
from app.models.conversation import Conversation, ConversationMessage
|
|
from app.models.flashcard import Flashcard, FlashcardDeck
|
|
from app.models.question import Question
|
|
from app.routers import ai_mode
|
|
from app.services import ai_mode_service
|
|
|
|
|
|
class CitationContractTests(unittest.TestCase):
|
|
"""The step that makes a hallucinated citation impossible rather than rare."""
|
|
|
|
def sources(self):
|
|
return [
|
|
{"kind": "article", "ref": "7", "id": 7, "title": "Febrile seizures"},
|
|
{"kind": "section", "ref": "7#abc", "id": 7, "section_id": "abc", "title": "Workup"},
|
|
]
|
|
|
|
def test_a_citation_that_was_not_retrieved_is_deleted(self):
|
|
reply, citations = ai_mode_service.enforce_citations(
|
|
"Fever first [[article:7]]. Then lumbar puncture [[article:99]].", self.sources())
|
|
self.assertIn("[[article:7]]", reply)
|
|
self.assertNotIn("99", reply)
|
|
self.assertEqual([c["id"] for c in citations], [7])
|
|
|
|
def test_deleting_a_marker_does_not_leave_broken_punctuation(self):
|
|
reply, _ = ai_mode_service.enforce_citations(
|
|
"This is true [[article:404]].", self.sources())
|
|
self.assertEqual(reply, "This is true.")
|
|
|
|
def test_a_url_the_model_invents_is_not_a_citation(self):
|
|
# Only the marker form counts, so a plausible-looking link cannot smuggle
|
|
# itself into the citation list.
|
|
reply, citations = ai_mode_service.enforce_citations(
|
|
"See https://uptodate.com/febrile-seizures for more.", self.sources())
|
|
self.assertEqual(citations, [])
|
|
self.assertIn("uptodate.com", reply) # left in the prose, cited by nothing
|
|
|
|
def test_the_same_source_cited_twice_is_listed_once(self):
|
|
_, citations = ai_mode_service.enforce_citations(
|
|
"One [[article:7]]. Two [[article:7]].", self.sources())
|
|
self.assertEqual(len(citations), 1)
|
|
|
|
def test_a_section_citation_keeps_the_section_it_points_at(self):
|
|
_, citations = ai_mode_service.enforce_citations("Here [[section:7#abc]].", self.sources())
|
|
self.assertEqual(citations[0]["section_id"], "abc")
|
|
|
|
def test_with_nothing_close_it_says_so_and_then_helps(self):
|
|
"""Refusing outright reads as a broken assistant, not a careful one.
|
|
|
|
The old behaviour was to say nothing matched and stop. It is honest to
|
|
name the gap; it is not honest to pretend an unrelated shortlist
|
|
supports the answer, and it is not useful to withhold one entirely.
|
|
"""
|
|
prompt = ai_mode_service.build_prompt([], "open")
|
|
self.assertIn("Nothing in this learner's library covers their question", prompt)
|
|
self.assertIn("answer from general knowledge", prompt)
|
|
self.assertIn("Do not cite anything", prompt)
|
|
# And nothing it writes can be cited anyway.
|
|
reply, citations = ai_mode_service.enforce_citations("Anything [[article:1]].", [])
|
|
self.assertEqual(citations, [])
|
|
self.assertEqual(reply, "Anything.")
|
|
|
|
def test_something_adjacent_is_named_as_adjacent(self):
|
|
# `sources()` carries only what citation enforcement needs; a prompt
|
|
# also prints the text of each source.
|
|
with_text = [{**s, "text": "Body"} for s in self.sources()]
|
|
prompt = ai_mode_service.build_prompt(with_text, "adjacent")
|
|
self.assertIn("closest things", prompt)
|
|
self.assertIn("[[section:7#abc]]", prompt) # still citable
|
|
|
|
def test_the_three_states_are_chosen_by_the_number_not_the_model(self):
|
|
sources = self.sources()
|
|
self.assertEqual(ai_mode_service.answer_mode(0.72, sources), "sourced")
|
|
self.assertEqual(ai_mode_service.answer_mode(0.52, sources), "adjacent")
|
|
# 0.49 is where "discuss love" and "photosynthesis" land against this
|
|
# corpus, alongside "tell me a joke" — noise, not adjacency.
|
|
self.assertEqual(ai_mode_service.answer_mode(0.49, sources), "open")
|
|
self.assertEqual(ai_mode_service.answer_mode(0.90, []), "open")
|
|
# Unmeasurable is not low: retrieval found these by other means, and
|
|
# dropping every citation because the ruler is missing would be worse.
|
|
self.assertEqual(ai_mode_service.answer_mode(None, sources), "sourced")
|
|
|
|
def test_a_greeting_is_not_a_query(self):
|
|
"""Whatever it scores, "hello" is not a question about anything.
|
|
|
|
This is the bug in the screenshot: a greeting came back with four
|
|
confident citations, because reciprocal-rank fusion always returns an
|
|
order and "ok" happens to embed at 0.51 against a corpus of clinical
|
|
prose — the wrong side of the adjacency threshold.
|
|
"""
|
|
sources = self.sources()
|
|
for greeting in ["hi", "Hello", "hello!", " thanks ", "ok thanks",
|
|
"good morning", "Thank you!", "bye", "",
|
|
"who are you?", "what can you do"]:
|
|
self.assertTrue(ai_mode_service.is_small_talk(greeting), greeting)
|
|
self.assertEqual(
|
|
ai_mode_service.answer_mode(0.9, sources, greeting), "chat", greeting)
|
|
|
|
def test_a_short_clinical_question_is_still_a_query(self):
|
|
# The guard is a closed vocabulary rather than a length rule, precisely
|
|
# so that these keep reaching retrieval.
|
|
sources = self.sources()
|
|
for query in ["croup dose?", "ok to give ibuprofen at 3 months?",
|
|
"no stridor now what", "hi flow nasal cannula",
|
|
"thanks to which vaccine has Hib fallen?"]:
|
|
self.assertFalse(ai_mode_service.is_small_talk(query), query)
|
|
self.assertEqual(
|
|
ai_mode_service.answer_mode(0.72, sources, query), "sourced", query)
|
|
|
|
def test_the_chat_prompt_offers_nothing_and_claims_nothing(self):
|
|
prompt = ai_mode_service.build_prompt(self.sources(), "chat")
|
|
# No shortlist reaches a turn that was never a query, so there is
|
|
# nothing for the model to cite even if it tries.
|
|
self.assertNotIn("[[section:7#abc]]", prompt)
|
|
self.assertNotIn("SOURCES", prompt)
|
|
# And unlike "open", it does not announce a gap nobody asked about.
|
|
self.assertNotIn("Nothing in this learner's library", prompt)
|
|
|
|
|
|
class _AiModeBase(unittest.TestCase):
|
|
"""Fixtures shared by the route and retrieval cases; holds no tests itself."""
|
|
|
|
def setUp(self):
|
|
# The daily limit is enforced through the real Redis, so without this a
|
|
# suite that passes today fails once it has been run sixty times — the
|
|
# tests would be measuring shared state rather than this code.
|
|
self._no_limit = patch("app.routers.ai_mode.check_rate_limit", lambda **kwargs: None)
|
|
self._no_limit.start()
|
|
self.addCleanup(self._no_limit.stop)
|
|
|
|
self.bank = fixtures.BuilderTests()
|
|
self.bank.setUp()
|
|
self.client = self.bank.client
|
|
self.client.app.include_router(ai_mode.router, prefix='/ai')
|
|
self.db = self.bank.db
|
|
|
|
self.db.add(Article(id=7, slug='febrile-seizures', title='Febrile seizures',
|
|
summary='A seizure with fever in a young child',
|
|
sections=[{"id": "a" * 32, "slug": "workup", "title": "Workup", "content": "LP"}],
|
|
status='published', user_id=3))
|
|
self.db.add(ArticleSectionIndex(id=1, article_id=7, section_id='a' * 32,
|
|
title='Workup', content='Lumbar puncture in a febrile infant'))
|
|
self.db.add(FlashcardDeck(id=1, title='Mine', user_id=1))
|
|
self.db.flush()
|
|
self.db.add(Flashcard(id=1, deck_id=1, front='Febrile seizure duration', back='Under 15 minutes'))
|
|
self.db.query(Question).filter(Question.id == 1).update(
|
|
{"question_text": "A child with a febrile seizure lasting two minutes"})
|
|
self.db.commit()
|
|
self.bank.user = self.bank.owner
|
|
|
|
def tearDown(self):
|
|
self.bank.tearDown()
|
|
|
|
def reply_with(self, text):
|
|
"""Stub the model, so the test is about the server's half of the contract."""
|
|
return patch("app.services.ai_service.achat",
|
|
new=AsyncMock(return_value=text))
|
|
|
|
def ask(self, message, conversation_id=None):
|
|
if conversation_id is None:
|
|
conversation_id = self.client.post('/ai/conversations').json()['id']
|
|
return conversation_id, self.client.post(
|
|
f'/ai/conversations/{conversation_id}/messages', json={'message': message})
|
|
|
|
|
|
class AiModeRouteTests(_AiModeBase):
|
|
def test_an_answer_keeps_only_the_citations_retrieval_found(self):
|
|
with self.reply_with("Yes [[article:7]], and also [[article:12345]]."):
|
|
_, response = self.ask('febrile seizure')
|
|
self.assertEqual(response.status_code, 200, response.text)
|
|
body = response.json()['message']
|
|
self.assertNotIn('12345', body['content'])
|
|
self.assertEqual([c['id'] for c in body['citations']], [7])
|
|
|
|
def test_the_first_question_names_the_thread(self):
|
|
with self.reply_with("An answer."):
|
|
conversation_id, response = self.ask('What causes febrile seizures?')
|
|
self.assertEqual(response.json()['title'], 'What causes febrile seizures?')
|
|
# And a later message does not rename it out from under the learner.
|
|
with self.reply_with("Another answer."):
|
|
_, second = self.ask('And the workup?', conversation_id)
|
|
self.assertEqual(second.json()['title'], 'What causes febrile seizures?')
|
|
|
|
def test_a_greeting_does_not_name_the_thread(self):
|
|
"""A rail of threads called "hi" is worse than one called New chat.
|
|
|
|
The name waits for the turn the conversation turns out to be about,
|
|
which is usually the very next one.
|
|
"""
|
|
with self.reply_with("Hello! What would you like to work on?"):
|
|
conversation_id, first = self.ask('hi')
|
|
self.assertEqual(first.json()['title'], 'New chat')
|
|
self.assertEqual(first.json()['source_count'], 0)
|
|
|
|
with self.reply_with("Fever, usually."):
|
|
_, second = self.ask('What causes febrile seizures?', conversation_id)
|
|
self.assertEqual(second.json()['title'], 'What causes febrile seizures?')
|
|
|
|
def test_both_turns_are_stored_so_a_thread_can_be_reopened(self):
|
|
with self.reply_with("Because of fever [[article:7]]."):
|
|
conversation_id, _ = self.ask('why do febrile seizures happen?')
|
|
messages = self.client.get(f'/ai/conversations/{conversation_id}').json()['messages']
|
|
self.assertEqual([m['role'] for m in messages], ['user', 'assistant'])
|
|
# The citations are stored as they were filtered, not recomputed later.
|
|
self.assertEqual(messages[1]['citations'][0]['id'], 7)
|
|
self.assertEqual(messages[0]['citations'], [])
|
|
|
|
def test_a_thread_belongs_to_one_learner(self):
|
|
with self.reply_with("An answer."):
|
|
conversation_id, _ = self.ask('febrile seizure')
|
|
self.bank.user = self.bank.peer
|
|
# Not 403: whether somebody else's thread exists is not this user's business.
|
|
self.assertEqual(self.client.get(f'/ai/conversations/{conversation_id}').status_code, 404)
|
|
self.assertEqual(self.client.delete(f'/ai/conversations/{conversation_id}').status_code, 404)
|
|
self.assertEqual(self.client.post(
|
|
f'/ai/conversations/{conversation_id}/messages', json={'message': 'hello'}).status_code, 404)
|
|
self.assertEqual(self.client.get('/ai/conversations').json(), [])
|
|
|
|
def test_a_model_failure_is_reported_not_stored(self):
|
|
conversation_id = self.client.post('/ai/conversations').json()['id']
|
|
with patch("app.services.ai_service.achat",
|
|
new=AsyncMock(side_effect=RuntimeError("down"))):
|
|
response = self.client.post(f'/ai/conversations/{conversation_id}/messages',
|
|
json={'message': 'febrile seizure'})
|
|
self.assertEqual(response.status_code, 502)
|
|
# A half-written exchange is worse than none: the question is not kept.
|
|
self.assertEqual(self.db.query(ConversationMessage).count(), 0)
|
|
|
|
def test_an_empty_answer_is_refused_rather_than_drawn_as_a_blank_card(self):
|
|
conversation_id = self.client.post('/ai/conversations').json()['id']
|
|
with self.reply_with(" "):
|
|
response = self.client.post(f'/ai/conversations/{conversation_id}/messages',
|
|
json={'message': 'febrile seizure'})
|
|
self.assertEqual(response.status_code, 502)
|
|
self.assertEqual(self.db.query(ConversationMessage).count(), 0)
|
|
|
|
def test_an_answer_that_was_only_an_invented_citation_is_refused(self):
|
|
conversation_id = self.client.post('/ai/conversations').json()['id']
|
|
# Every word of it goes when the marker nobody can vouch for goes.
|
|
with self.reply_with("[[article:9999]]"):
|
|
response = self.client.post(f'/ai/conversations/{conversation_id}/messages',
|
|
json={'message': 'febrile seizure'})
|
|
self.assertEqual(response.status_code, 502)
|
|
self.assertEqual(self.db.query(ConversationMessage).count(), 0)
|
|
|
|
def test_a_thread_is_named_tidily_from_its_first_question(self):
|
|
conversation_id = self.client.post('/ai/conversations').json()['id']
|
|
with self.reply_with("An answer."):
|
|
response = self.client.post(
|
|
f'/ai/conversations/{conversation_id}/messages',
|
|
json={'message': 'hi, how do i treat a febrile seizure?'})
|
|
# Not the learner's typing verbatim: no opener, a capital, and "I".
|
|
self.assertEqual(response.json()['title'], 'How do I treat a febrile seizure?')
|
|
|
|
def test_deleting_a_thread_takes_its_messages(self):
|
|
with self.reply_with("An answer."):
|
|
conversation_id, _ = self.ask('febrile seizure')
|
|
self.assertEqual(self.client.delete(f'/ai/conversations/{conversation_id}').status_code, 204)
|
|
self.assertEqual(self.db.query(Conversation).count(), 0)
|
|
self.assertEqual(self.db.query(ConversationMessage).count(), 0)
|
|
|
|
|
|
class RetrievalTests(_AiModeBase):
|
|
def test_retrieval_offers_only_what_this_learner_may_see(self):
|
|
self.db.add(Article(id=8, slug='draft-febrile', title='Febrile draft',
|
|
summary='Unpublished febrile notes', sections=[],
|
|
status='draft', user_id=3))
|
|
self.db.commit()
|
|
|
|
self.bank.user = self.bank.owner
|
|
found = ai_mode_service.retrieve(self.db, self.bank.owner, 'febrile')
|
|
self.assertNotIn(8, [s['id'] for s in found if s['kind'] == 'article'])
|
|
# Question 3 is another user's private question.
|
|
self.assertNotIn(3, [s['id'] for s in found if s['kind'] == 'question'])
|
|
|
|
self.bank.user = self.bank.mod
|
|
found = ai_mode_service.retrieve(self.db, self.bank.mod, 'febrile')
|
|
self.assertIn(8, [s['id'] for s in found if s['kind'] == 'article'])
|
|
|
|
def test_a_question_source_carries_the_stem_and_not_the_answer(self):
|
|
found = ai_mode_service.retrieve(self.db, self.bank.owner, 'febrile seizure')
|
|
questions = [s for s in found if s['kind'] == 'question']
|
|
self.assertTrue(questions)
|
|
for source in questions:
|
|
self.assertNotIn('Full explanation', source['text'])
|
|
self.assertNotIn('yes', source['text'].split())
|
|
|
|
def test_a_curated_link_between_two_hits_lifts_both(self):
|
|
self.db.add(QuestionArticleLink(question_id=1, article_id=7, section_id=None))
|
|
self.db.commit()
|
|
found = ai_mode_service.retrieve(self.db, self.bank.owner, 'febrile seizure')
|
|
curated = [s for s in found if s.get('curated')]
|
|
# An educator tied these two together; both surfacing for one query is
|
|
# evidence rather than coincidence.
|
|
self.assertIn(('article', 7), [(s['kind'], s['id']) for s in curated])
|
|
self.assertIn(('question', 1), [(s['kind'], s['id']) for s in curated])
|
|
|
|
def test_the_prompt_carries_the_shortlist_and_the_rules(self):
|
|
found = ai_mode_service.retrieve(self.db, self.bank.owner, 'febrile')
|
|
prompt = ai_mode_service.build_prompt(found)
|
|
self.assertIn('[[article:7]]', prompt)
|
|
self.assertIn('never cite a marker that is not listed here', prompt)
|
|
self.assertIn('Never reveal the answer to a practice question', prompt)
|
|
# Asked for five questions, it used to account for itself: how many it
|
|
# had seen, what it might go and fetch. The count is not the learner's
|
|
# to set and not the model's to discuss.
|
|
self.assertIn('Ignore', prompt)
|
|
self.assertIn('never how many questions there are', prompt)
|
|
|
|
|
|
class DeterminismTests(unittest.TestCase):
|
|
"""The same question, the same answer.
|
|
|
|
Measured over fifteen runs before any of this: retrieval was already
|
|
deterministic — identical shortlist and scores every time, and the citation
|
|
checker stripped none of the 45 markers written. Generation was the whole
|
|
variance. At temperature 0.3 the same sources and the same prompt produced
|
|
answers differing by 15-70% of their text; one differential swung between a
|
|
35-word uncited paraphrase and a 180-word cited list. Temperature 0 alone
|
|
was not enough (three runs still differed); temperature 0 with a fixed seed
|
|
came back byte-identical.
|
|
"""
|
|
|
|
def test_the_seed_follows_the_question_not_the_clock(self):
|
|
import hashlib
|
|
|
|
def seed_for(question):
|
|
return int(hashlib.sha256(
|
|
" ".join(question.lower().split()).encode()).hexdigest()[:8], 16)
|
|
|
|
# Same question, same seed — including through casing and spacing, so
|
|
# two people who type it differently still get one answer.
|
|
self.assertEqual(seed_for("What causes croup?"),
|
|
seed_for(" what causes CROUP? "))
|
|
# A different question is not pinned to the same sample.
|
|
self.assertNotEqual(seed_for("What causes croup?"),
|
|
seed_for("What causes bronchiolitis?"))
|
|
|
|
def test_a_sourced_answer_must_cite_something(self):
|
|
from app.services.ai_mode_service import build_prompt
|
|
prompt = build_prompt([{
|
|
"kind": "article", "ref": "7", "title": "Croup",
|
|
"text": "Croup is …", "score": 1.0,
|
|
}])
|
|
self.assertIn("at least one sentence must carry a citation", prompt)
|
|
|
|
def test_an_open_answer_is_not_asked_to_cite(self):
|
|
# Nothing to cite, so the instruction would be an invitation to invent.
|
|
from app.services.ai_mode_service import build_prompt
|
|
self.assertNotIn("must carry a citation", build_prompt([], mode="open"))
|
|
self.assertNotIn("must carry a citation", build_prompt([], mode="chat"))
|