File size: 12,020 Bytes
a5a31c9
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
 
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
"""End-to-end HTTP behaviour, with the model and vector store stubbed out.

These tests are where the security claims meet the wire: that the keyless path
really is keyless, that the LLM is never reached below the threshold, and that
a credential cannot come back out in a response body.
"""

from __future__ import annotations

import numpy as np
import pytest
from fastapi.testclient import TestClient

from ehekim import api, llm
from ehekim.config import MODEL_REFUSAL_MESSAGE_TR, REFUSAL_MESSAGE_TR
from ehekim.vectorstore import SearchHit

GOOD_KEY = "sk-abcdef0123456789abcdef0123456789"


def make_hit(similarity: float, chunk_id: str = "c1") -> SearchHit:
    return SearchHit(
        chunk_id=chunk_id,
        chunk_text="Migren, zonklayıcı baş ağrısıdır.",
        similarity=similarity,
        url="https://hastane.test/migren",
        title="Migren Nedir?",
        source="acibadem",
        parent_id="p1",
        chunk_index=0,
    )


class FakeEmbedder:
    device = "cpu"

    def encode_query(self, query: str) -> np.ndarray:
        return np.ones(8, dtype=np.float32)


class FakeStore:
    def __init__(self, hits: list[SearchHit]) -> None:
        self.hits = hits

    def count(self) -> int:
        return len(self.hits)

    def query(self, embedding, top_k: int) -> list[SearchHit]:
        return list(self.hits[:top_k])

    def get_siblings(self, parent_id: str, indices) -> list[SearchHit]:
        wanted = set(indices)
        return [
            h for h in self.hits
            if h.parent_id == parent_id and h.chunk_index in wanted
        ]


@pytest.fixture
def client(monkeypatch) -> TestClient:
    """A client with app state injected directly, bypassing model loading."""
    monkeypatch.setattr(api.state, "embedder", FakeEmbedder(), raising=False)
    monkeypatch.setattr(
        api.state, "store", FakeStore([make_hit(0.82), make_hit(0.31, "c2")]), raising=False
    )
    # Not used as a context manager: lifespan (which loads the real model) is
    # deliberately not run.
    return TestClient(api.app)


class TestConfig:
    def test_exposes_catalogue_and_defaults(self, client):
        body = client.get("/api/config").json()
        assert body["embedding_dim"] == 768
        assert body["refusal_message"] == REFUSAL_MESSAGE_TR
        assert body["chunk_count"] == 2
        providers = {p["id"] for p in body["providers"]}
        assert "deepseek" in providers and "openrouter" in providers

    def test_catalogue_contains_no_credentials(self, client):
        """The catalogue ships key *hints* ("sk-…") but never a real key."""
        from ehekim.security import contains_secret

        raw = client.get("/api/config").text
        assert contains_secret(raw) is False

    def test_offers_more_than_one_model_family(self, client):
        body = client.get("/api/config").json()
        families = {m["family"] for p in body["providers"] for m in p["models"]}
        assert len(families) >= 3


class TestSecurityHeaders:
    def test_sets_hardening_headers(self, client):
        response = client.get("/api/health")
        assert "default-src 'self'" in response.headers["content-security-policy"]
        assert response.headers["x-content-type-options"] == "nosniff"
        assert response.headers["x-frame-options"] == "DENY"
        assert response.headers["referrer-policy"] == "no-referrer"
        assert response.headers["cache-control"] == "no-store"

    def test_rejects_oversized_bodies(self, client):
        response = client.post(
            "/api/search",
            content=b"x" * (api.MAX_BODY_BYTES + 1),
            headers={"Content-Type": "application/json", "Content-Length": str(api.MAX_BODY_BYTES + 1)},
        )
        assert response.status_code == 413


class TestSemanticSearch:
    def test_returns_similarities_and_partitions_on_the_threshold(self, client):
        body = client.post("/api/search", json={"query": "migren", "top_k": 5, "threshold": 0.55}).json()
        assert body["mode"] == "search"
        assert body["grounded"] is True
        assert body["best_similarity"] == pytest.approx(0.82)
        passed = [r for r in body["results"] if r["passed_threshold"]]
        rejected = [r for r in body["results"] if not r["passed_threshold"]]
        assert len(passed) == 1 and len(rejected) == 1
        assert passed[0]["similarity"] == pytest.approx(0.82)

    def test_needs_no_api_key(self, client):
        assert client.post("/api/search", json={"query": "migren"}).status_code == 200

    def test_high_threshold_yields_the_refusal_notice(self, client):
        body = client.post("/api/search", json={"query": "migren", "threshold": 0.95}).json()
        assert body["grounded"] is False
        assert body["notice"] == REFUSAL_MESSAGE_TR

    @pytest.mark.parametrize(
        "payload",
        [
            {"query": ""},
            {"query": "x", "top_k": 0},
            {"query": "x", "top_k": 999},
            {"query": "x", "threshold": 1.5},
            {"query": "x", "threshold": -0.1},
            {"query": "x", "unexpected": "field"},
        ],
    )
    def test_rejects_invalid_payloads(self, client, payload):
        assert client.post("/api/search", json=payload).status_code == 422


class TestRagAuth:
    def test_missing_key_is_rejected(self, client):
        assert client.post("/api/ask", json={"query": "migren"}).status_code == 401

    def test_malformed_key_is_rejected(self, client):
        response = client.post(
            "/api/ask", json={"query": "migren"}, headers={"X-Provider-Key": "short"}
        )
        assert response.status_code == 401

    def test_rejection_message_does_not_echo_the_key(self, client):
        secret = "sk-thisisaverysecretkey0123456789"
        response = client.post(
            "/api/ask", json={"query": "migren"}, headers={"X-Provider-Key": secret + " bad"}
        )
        assert secret not in response.text


class TestThresholdGateBlocksTheModel:
    def test_below_threshold_the_llm_is_never_called(self, client, monkeypatch):
        def explode(**kwargs):
            raise AssertionError("Eşiğin altında LLM çağrılmamalıydı.")

        monkeypatch.setattr(llm, "generate", explode)
        body = client.post(
            "/api/ask",
            json={"query": "ay'a nasıl gidilir", "threshold": 0.95},
            headers={"X-Provider-Key": GOOD_KEY},
        ).json()
        assert body["refused"] is True
        assert body["grounded"] is False
        assert body["answer"] == REFUSAL_MESSAGE_TR
        assert body["refusal_reason"] == "below_threshold"
        assert body["model"] is None

    def test_model_refusal_is_reported_as_the_second_layer(self, client, monkeypatch):
        """Passages cleared the gate, but the model said the answer is not there."""
        monkeypatch.setattr(
            llm,
            "generate",
            lambda **kw: llm.LLMResult(content=MODEL_REFUSAL_MESSAGE_TR, model="deepseek-v4-flash"),
        )
        body = client.post(
            "/api/ask",
            json={"query": "Hodgkin lenfomada sağkalım oranı nedir?", "threshold": 0.5},
            headers={"X-Provider-Key": GOOD_KEY},
        ).json()
        assert body["grounded"] is True          # retrieval did find close passages
        assert body["refused"] is True           # but the model declined
        assert body["refusal_reason"] == "model_insufficient_context"
        assert body["answer"] == MODEL_REFUSAL_MESSAGE_TR

    def test_empty_completion_becomes_a_refusal_not_an_empty_answer(self, client, monkeypatch):
        monkeypatch.setattr(
            llm, "generate", lambda **kw: llm.LLMResult(content="", model="m")
        )
        body = client.post(
            "/api/ask", json={"query": "migren", "threshold": 0.5},
            headers={"X-Provider-Key": GOOD_KEY},
        ).json()
        assert body["refused"] is True
        assert body["answer"] == MODEL_REFUSAL_MESSAGE_TR


class TestRagSuccess:
    def test_answers_and_forwards_the_key_without_returning_it(self, client, monkeypatch):
        seen = {}

        def fake_generate(*, model_key, api_key, messages, timeout, **kwargs):
            seen["model_key"] = model_key
            seen["api_key"] = api_key
            return llm.LLMResult(
                content="Migren zonklayıcı bir baş ağrısıdır [1].",
                model="deepseek-v4-flash",
                reasoning="gizli düşünce",
                prompt_tokens=100,
                completion_tokens=20,
                reasoning_tokens=5,
            )

        monkeypatch.setattr(llm, "generate", fake_generate)
        response = client.post(
            "/api/ask",
            json={"query": "migren nedir", "threshold": 0.5},
            headers={"X-Provider-Key": GOOD_KEY},
        )
        body = response.json()

        assert seen["api_key"] == GOOD_KEY          # forwarded upstream
        assert GOOD_KEY not in response.text        # never returned to the client
        assert body["refused"] is False
        assert body["answer"].startswith("Migren")
        assert body["usage"]["reasoning_tokens"] == 5

    def test_reasoning_is_withheld_unless_requested(self, client, monkeypatch):
        monkeypatch.setattr(
            llm,
            "generate",
            lambda **kw: llm.LLMResult(content="yanıt", model="m", reasoning="gizli"),
        )
        off = client.post(
            "/api/ask", json={"query": "migren", "threshold": 0.5},
            headers={"X-Provider-Key": GOOD_KEY},
        ).json()
        assert off["reasoning"] is None

        on = client.post(
            "/api/ask",
            json={"query": "migren", "threshold": 0.5, "include_reasoning": True},
            headers={"X-Provider-Key": GOOD_KEY},
        ).json()
        assert on["reasoning"] == "gizli"

    def test_upstream_errors_are_sanitised_before_relay(self, client, monkeypatch):
        def failing(**kwargs):
            raise llm.LLMError(
                f"Authentication Fails, Your api key: {GOOD_KEY} is invalid", status_code=401
            )

        monkeypatch.setattr(llm, "generate", failing)
        response = client.post(
            "/api/ask", json={"query": "migren", "threshold": 0.5},
            headers={"X-Provider-Key": GOOD_KEY},
        )
        assert response.status_code == 401
        assert GOOD_KEY not in response.text
        assert "REDACTED" in response.text


class TestLogRedaction:
    """The scrubber must clean strings without breaking record formatting."""

    def _record(self, msg, args):
        import logging

        return logging.LogRecord("uvicorn.access", logging.INFO, __file__, 1, msg, args, None)

    def test_scrubs_credentials_from_message_args(self):
        record = self._record("key=%s", (f"Bearer {GOOD_KEY}",))
        api.RedactingFilter().filter(record)
        assert GOOD_KEY not in record.getMessage()
        assert "REDACTED" in record.getMessage()

    def test_preserves_numeric_args_so_uvicorn_access_log_still_formats(self):
        """Regression: coercing args to str made '%d' raise on the status code."""
        record = self._record(
            '%s - "%s %s HTTP/%s" %d', ("127.0.0.1:1234", "POST", "/api/ask", "1.1", 413)
        )
        api.RedactingFilter().filter(record)
        assert record.getMessage() == '127.0.0.1:1234 - "POST /api/ask HTTP/1.1" 413'

    def test_handles_dict_args(self):
        record = self._record("%(k)s %(n)d", {"k": f"sk-{'a1b2c3d4' * 4}", "n": 7})
        api.RedactingFilter().filter(record)
        assert record.args["n"] == 7


class TestModelCatalogue:
    def test_unknown_model_key_is_a_client_error(self):
        with pytest.raises(llm.LLMError) as info:
            llm.get_model_spec("bogus:model")
        assert info.value.status_code == 400

    def test_deepseek_default_enables_thinking(self):
        spec = llm.get_model_spec(None)
        assert spec.model == "deepseek-v4-flash"
        assert spec.supports_thinking is True