pastor-online/backend/tests/test_llm.py

"""Testes dos adaptadores de LLM.

Cada adaptador converte o formato de streaming do próprio provedor em deltas de
texto. Os testes usam transportes falsos: verificam a conversão, não a rede.
"""

from __future__ import annotations

import json
import sys
from pathlib import Path

import httpx
import pytest

sys.path.insert(0, str(Path(__file__).resolve().parents[1]))

from app.config import Settings  # noqa: E402
from app.llm import (  # noqa: E402
    SYSTEM_PROMPT,
    AnthropicProvider,
    LLMProviderError,
    OllamaProvider,
    OpenAIProvider,
    StubProvider,
    build_provider,
)


class FakeStreamResponse:
    """Dublê do contexto assíncrono devolvido por `httpx.AsyncClient.stream`."""

    status_code = 200

    def __init__(self, chunks):
        self._chunks = chunks

    async def __aenter__(self):
        return self

    async def __aexit__(self, *exc_info):
        return False

    async def aiter_lines(self):
        for chunk in self._chunks:
            text = chunk.decode() if isinstance(chunk, bytes) else chunk
            for line in text.splitlines():
                yield line

    async def aread(self):
        return b"erro"


class FakeStreamClient:
    """Dublê do `httpx.AsyncClient` que devolve um stream fixo."""

    def __init__(self, chunks):
        self._chunks = chunks

    def stream(self, *args, **kwargs):
        return FakeStreamResponse(self._chunks)


def collect(provider, messages=None, system=SYSTEM_PROMPT):
    import asyncio

    async def run():
        chunks = []
        async for delta in provider.stream(messages or [{"role": "user", "content": "olá"}], system):
            chunks.append(delta)
        return "".join(chunks)

    return asyncio.run(run())


# --------------------------------------------------------------------------- #
# Seleção de provedor
# --------------------------------------------------------------------------- #

def test_build_provider_selects_ollama():
    provider = build_provider(Settings(llm_provider="ollama"))
    assert isinstance(provider, OllamaProvider)


def test_build_provider_selects_openai():
    provider = build_provider(Settings(llm_provider="openai", openai_api_key="sk-x"))
    assert isinstance(provider, OpenAIProvider)


def test_build_provider_selects_groq_as_openai_compatible():
    provider = build_provider(Settings(llm_provider="groq", groq_api_key="gsk-x"))
    assert isinstance(provider, OpenAIProvider)
    assert provider.base_url.startswith("https://api.groq.com")


def test_build_provider_selects_anthropic():
    provider = build_provider(Settings(llm_provider="anthropic", anthropic_api_key="sk-x"))
    assert isinstance(provider, AnthropicProvider)


def test_build_provider_selects_stub():
    provider = build_provider(Settings(llm_provider="stub"))
    assert isinstance(provider, StubProvider)


def test_build_provider_rejects_unknown_provider():
    with pytest.raises(ValueError, match="não suportado"):
        build_provider(Settings(llm_provider="desconhecido"))


def test_openai_provider_requires_api_key():
    with pytest.raises(ValueError, match="OPENAI_API_KEY"):
        build_provider(Settings(llm_provider="openai", openai_api_key=""))


# --------------------------------------------------------------------------- #
# System prompt
# --------------------------------------------------------------------------- #

def test_system_prompt_contains_the_spec_directives():
    assert "conselheiro espiritual e Pastor Online" in SYSTEM_PROMPT
    assert "única fonte de autoridade" in SYSTEM_PROMPT
    assert "livro, capítulo e versículo" in SYSTEM_PROMPT
    assert "tom acolhedor" in SYSTEM_PROMPT
    assert "Não crie doutrina própria" in SYSTEM_PROMPT


def test_system_prompt_contains_the_crisis_scope():
    assert "emergência" in SYSTEM_PROMPT.lower()
    assert "diagnóstico" in SYSTEM_PROMPT.lower() or "diagnost" in SYSTEM_PROMPT.lower()


# --------------------------------------------------------------------------- #
# Conversão de streaming
# --------------------------------------------------------------------------- #

def test_ollama_provider_converts_ndjson_stream(monkeypatch):
    provider = OllamaProvider(base_url="http://ollama:11434", model="qwen2.5:1.5b")

    lines = [
        {"message": {"content": "Olá "}, "done": False},
        {"message": {"content": "mundo"}, "done": False},
        {"message": {"content": ""}, "done": True},
    ]

    async def fake_stream(*args, **kwargs):
        for line in lines:
            yield f"{json.dumps(line)}\n".encode()

    monkeypatch.setattr(
        provider, "_client", FakeStreamClient([f"{json.dumps(line)}\n" for line in lines])
    )

    assert collect(provider) == "Olá mundo"


def test_openai_provider_converts_sse_stream(monkeypatch):
    provider = OpenAIProvider(
        base_url="https://api.openai.com/v1", model="gpt-4o-mini", api_key="sk-test"
    )

    chunks = [
        'data: {"choices":[{"delta":{"content":"Olá "}}]}\n\n'.encode(),
        'data: {"choices":[{"delta":{"content":"mundo"}}]}\n\n'.encode(),
        b"data: [DONE]\n\n",
    ]

    async def fake_stream(*args, **kwargs):
        for chunk in chunks:
            yield chunk

    monkeypatch.setattr(provider, "_client", FakeStreamClient(chunks))

    assert collect(provider) == "Olá mundo"


def test_anthropic_provider_converts_event_stream(monkeypatch):
    provider = AnthropicProvider(
        base_url="https://api.anthropic.com", model="claude-sonnet-4-5", api_key="sk-test"
    )

    events = [
        'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"type":"text_delta","text":"Olá "}}\n\n'.encode(),
        'event: content_block_delta\ndata: {"type":"content_block_delta","delta":{"type":"text_delta","text":"mundo"}}\n\n'.encode(),
    ]

    async def fake_stream(*args, **kwargs):
        for event in events:
            yield event

    monkeypatch.setattr(provider, "_client", FakeStreamClient(events))

    assert collect(provider) == "Olá mundo"


def test_stub_provider_returns_deterministic_text():
    provider = StubProvider()
    text = collect(provider, [{"role": "user", "content": "Quem criou o mundo?"}])
    assert "Gênesis 1:1" in text


# --------------------------------------------------------------------------- #
# Erros
# --------------------------------------------------------------------------- #

def test_provider_error_does_not_leak_the_api_key():
    provider = OpenAIProvider(
        base_url="https://api.openai.com/v1", model="gpt-4o-mini", api_key="sk-super-secreta"
    )

    def raise_error(*args, **kwargs):
        raise httpx.ConnectError("connection refused")

    monkeypatched = provider._client
    try:
        provider._client = type(
            "FailingClient",
            (),
            {"stream": staticmethod(raise_error)},
        )()
        with pytest.raises(LLMProviderError) as exc:
            collect(provider)
        assert "sk-super-secreta" not in str(exc.value)
        assert "sk-super-secreta" not in str(exc.value.__cause__ or "")
    finally:
        provider._client = monkeypatched


def test_system_prompt_declares_the_off_topic_scope():
    """O portão por similaridade não funciona: precisa estar no prompt.

    Medido no corpus real, nenhum valor de MIN_SIMILARITY entre 0,30 e 0,50
    bloqueou uma única pergunta fora do tema — "como fazer um bolo de chocolate"
    bate 0,544 em Levítico 7:14 e "cotação do dólar" bate 0,494. Num corpus de
    31 mil versículos sempre existe algum acima de qualquer limiar utilizável,
    então a recusa tem de ser instruída ao modelo, não filtrada por nota.
    """
    lowered = SYSTEM_PROMPT.lower()

    assert "receita" in lowered
    assert "apenas do texto bíblico" in lowered
    assert "não use os trechos recuperados" in lowered