"""Lesson pipeline: per-student cache (memory + Firestore), in-flight dedupe, concurrent video lookup, SSE stream, reasoner fallback, richness validation + repair, learner profile, prefetch.""" from __future__ import annotations import asyncio import json import os import threading import time from pathlib import Path from types import SimpleNamespace from typing import Callable, Dict, Iterator, List, Optional, Tuple import httpx import pytest from fastapi import FastAPI, Request from fastapi.testclient import TestClient os.environ.setdefault("ANTHROPIC_API_KEY", "mock-key-for-testing") from routes import rag_routes # noqa: E402 from services.ai_client import REASONER_MODEL as REASONER, AIRefusalError, Timeout # noqa: E402 CHUNKS = [ { "content": "A rational function is a quotient of two polynomial functions.", "source_file": "SHS_GM_Q1.pdf", "score": 0.95, "content_domain": "general", } ] PAYLOAD = { "topic": "Rational Functions", "subject": "General Mathematics", "quarter": 1, "lessonTitle": "Introduction to Rational Functions", "lessonId": "gm-q1-rational", } VIDEOS = [{"videoId": "abc123", "title": "Rational Functions", "channelTitle": "Math PH", "thumbnailUrl": "t.jpg"}] PROFILE = "Initial assessment: placed, starting quarter Q1\nWeak topics: Functions" SAMPLE_SSE_PATH = Path(__file__).resolve().parents[2] / ".unlazy" / "rag-lesson-stream.sample.sse" def _pad(text: str, length: int) -> str: return (text + " ") * (length // (len(text) + 1) + 1) def _lesson_dict(marker: str, rich: bool = True) -> Dict[str, object]: if not rich: return {"sections": [{"type": "introduction", "content": f"{marker} introduction."}]} return { "sections": [ {"type": "introduction", "title": "Introduction", "content": _pad(f"{marker} introduction.", 420)}, { "type": "key_concepts", "title": "Key Concepts", "content": _pad(f"{marker} concepts $f(x)=\\frac{{p(x)}}{{q(x)}}$.", 620), "callouts": [{"type": "warning", "text": "q(x) cannot be zero."}], }, {"type": "video", "title": "Video Lesson", "content": "Watch the video."}, { "type": "worked_examples", "title": "Worked Examples", "examples": [ {"problem": f"{marker} example {i}.", "steps": ["Step 1: a", "Step 2: b", "Step 3: c"], "answer": "x = 2"} for i in range(3) ], }, {"type": "important_notes", "title": "Important Notes", "bulletPoints": [f"{marker} note {i}." for i in range(4)]}, { "type": "try_it_yourself", "title": "Try It Yourself", "practiceProblems": [{"question": f"{marker} practice {i}.", "solution": "Solve step by step."} for i in range(5)], }, {"type": "summary", "title": "Summary", "content": _pad(f"{marker} summary.", 260)}, ] } def _lesson_json(marker: str, rich: bool = True) -> str: return json.dumps(_lesson_dict(marker, rich)) def _intro(response_json: Dict[str, object]) -> str: return str(response_json["sections"][0]["content"]) class Pipeline: """Test doubles for every external seam in rag_routes; counts calls.""" def __init__(self) -> None: self.stream_calls = 0 self.retrieval_calls = 0 self.inference_requests: List[SimpleNamespace] = [] self.stream_delay = 0.0 self.video_delay = 0.0 self.stream_outputs: List[str] = [] self.repair_output = _lesson_json("Repaired") self.stream_error: Optional[Exception] = None self.video_error: Optional[Exception] = None self.retrieval_error: Optional[Exception] = None self.chunks = CHUNKS self.verification: Dict[str, object] = {"verified": True, "pCorrect": 0.95} self.events: List[Tuple[str, float]] = [] self.store: Dict[Tuple[str, str], Dict[str, object]] = {} self.writes: List[Tuple[str, str, Dict[str, object]]] = [] self.profile: Callable[[str, rag_routes.RagLessonRequest], str] = lambda uid, payload: PROFILE def retrieve(self, **_kwargs: object) -> Tuple[List[Dict[str, object]], str]: self.retrieval_calls += 1 if self.retrieval_error is not None: raise self.retrieval_error return self.chunks, "exact_file" def stream( self, _messages: List[Dict[str, str]], model: str, report_progress: Callable[[str, int], None], _stop: threading.Event, ) -> str: assert model == REASONER self.stream_calls += 1 call_number = self.stream_calls self.events.append(("generate_start", time.perf_counter())) report_progress("thinking", 120) time.sleep(self.stream_delay) if self.stream_error is not None: raise self.stream_error report_progress("writing", 480) self.events.append(("generate_end", time.perf_counter())) if self.stream_outputs: return self.stream_outputs[min(call_number, len(self.stream_outputs)) - 1] return _lesson_json(f"Generated{call_number}") async def verify(self, reference_text: str, generated_text: str) -> Dict[str, object]: return self.verification def videos(self, *_args: object, **_kwargs: object) -> List[Dict[str, str]]: self.events.append(("video_start", time.perf_counter())) time.sleep(self.video_delay) if self.video_error is not None: raise self.video_error return VIDEOS def inference_client(self) -> SimpleNamespace: def generate_from_messages(request: SimpleNamespace) -> str: self.inference_requests.append(request) if request.max_new_tokens == 8192: return self.repair_output return _lesson_json("Fallback") return SimpleNamespace(generate_from_messages=generate_from_messages) def read_lesson(self, uid: str, key: str) -> Optional[Dict[str, object]]: return self.store.get((uid, key)) def write_lesson(self, uid: str, key: str, payload: object, lesson: Dict[str, object], profile_summary: str) -> None: self.writes.append((uid, key, lesson)) self.store[(uid, key)] = lesson @pytest.fixture def pipeline(monkeypatch: pytest.MonkeyPatch) -> Pipeline: fake = Pipeline() for store in (rag_routes._lesson_memory, rag_routes._retrieval_memory, rag_routes._lesson_jobs, rag_routes._prefetch_uids): store.clear() monkeypatch.setattr(rag_routes, "_prefetch_semaphore", asyncio.Semaphore(2)) monkeypatch.setattr(rag_routes, "retrieve_lesson_pdf_context", fake.retrieve) monkeypatch.setattr(rag_routes, "_stream_reasoner_lesson", fake.stream) monkeypatch.setattr(rag_routes, "verify_lesson_factuality", fake.verify) monkeypatch.setattr(rag_routes, "_fetch_youtube_videos", fake.videos) monkeypatch.setattr(rag_routes, "_log_rag_usage", lambda *args, **kwargs: None) monkeypatch.setattr(rag_routes, "_lesson_primary_model", lambda: REASONER) monkeypatch.setattr(rag_routes, "_get_inference_client", fake.inference_client) monkeypatch.setattr(rag_routes, "_read_student_lesson", fake.read_lesson) monkeypatch.setattr(rag_routes, "_write_student_lesson", fake.write_lesson) monkeypatch.setattr(rag_routes, "_load_learner_profile", lambda uid, payload: fake.profile(uid, payload)) # Keyword-capturing stand-in so the fallback contract (max_retries=1) is asserted without the real dataclass. monkeypatch.setattr(rag_routes, "InferenceRequest", SimpleNamespace) yield fake for store in (rag_routes._lesson_memory, rag_routes._retrieval_memory, rag_routes._lesson_jobs, rag_routes._prefetch_uids): store.clear() def _app() -> FastAPI: """Router under a stand-in for main.py AuthMiddleware: X-Test-Uid becomes request.state.user.""" app = FastAPI() @app.middleware("http") async def fake_auth(request: Request, call_next: Callable) -> object: uid = request.headers.get("X-Test-Uid") request.state.user = SimpleNamespace(uid=uid, role=request.headers.get("X-Test-Role", "student")) if uid else None return await call_next(request) app.include_router(rag_routes.router) return app def _as(uid: str) -> Dict[str, str]: return {"X-Test-Uid": uid} @pytest.fixture def client(pipeline: Pipeline) -> TestClient: with TestClient(_app()) as test_client: yield test_client def _key() -> str: return rag_routes._lesson_cache_key(rag_routes.RagLessonRequest(**PAYLOAD)) def wait_for_writes(pipeline: "Pipeline", count: int) -> None: """Firestore writes are fire-and-forget after the response; wait for them to land.""" deadline = time.monotonic() + 3.0 while len(pipeline.writes) < count and time.monotonic() < deadline: time.sleep(0.01) def parse_sse(body: str) -> List[Tuple[str, object]]: """Return [(event, data)] in order; comment lines become ("comment", text).""" events: List[Tuple[str, object]] = [] for block in body.replace("\r\n", "\n").split("\n\n"): if not block.strip(): continue event, data_lines = "message", [] for line in block.split("\n"): if line.startswith(":"): events.append(("comment", line[1:].strip())) elif line.startswith("event:"): event = line[len("event:"):].strip() elif line.startswith("data:"): data_lines.append(line[len("data:"):].strip()) if data_lines: events.append((event, json.loads("\n".join(data_lines)))) return events # ─── cache ───────────────────────────────────────────────────────────────────── def test_cache_serves_same_student_from_memory_without_regenerating(client: TestClient, pipeline: Pipeline) -> None: first = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) second = client.post("/api/rag/lesson", json={**PAYLOAD, "userId": "whatever"}, headers=_as("alice")) assert first.status_code == second.status_code == 200 assert pipeline.stream_calls == 1 assert second.json() == first.json() assert first.json()["activeModel"] == REASONER assert first.json()["personalized"] is True def test_cache_is_isolated_per_student(client: TestClient, pipeline: Pipeline) -> None: alice = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() bob = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("bob")).json() assert pipeline.stream_calls == 2 assert _intro(alice).startswith("Generated1") assert _intro(bob).startswith("Generated2") wait_for_writes(pipeline, 2) assert {(uid, key) for uid, key, _ in pipeline.writes} == {("alice", _key()), ("bob", _key())} def test_cache_payload_user_id_cannot_select_another_students_lesson(client: TestClient, pipeline: Pipeline) -> None: client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) mallory = client.post("/api/rag/lesson", json={**PAYLOAD, "userId": "alice"}, headers=_as("mallory")).json() assert pipeline.stream_calls == 2 assert _intro(mallory).startswith("Generated2") wait_for_writes(pipeline, 2) assert ("mallory", _key()) in pipeline.store def test_cache_anonymous_requests_are_not_cached_or_personalized(client: TestClient, pipeline: Pipeline) -> None: first = client.post("/api/rag/lesson", json=PAYLOAD).json() client.post("/api/rag/lesson", json=PAYLOAD) assert pipeline.stream_calls == 2 assert first["personalized"] is False assert pipeline.writes == [] def test_cache_firestore_hit_returns_without_generating(client: TestClient, pipeline: Pipeline) -> None: stored = {**_lesson_dict("Stored"), "activeModel": REASONER} pipeline.store[("alice", _key())] = stored plain = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) rag_routes._lesson_memory.clear() events = parse_sse(client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")).text) assert plain.json() == stored assert events == [("stage", {"stage": "cached"}), ("lesson", stored)] assert pipeline.stream_calls == 0 assert pipeline.retrieval_calls == 0 def test_cache_firestore_write_happens_after_success(client: TestClient, pipeline: Pipeline) -> None: response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() wait_for_writes(pipeline, 1) assert pipeline.writes == [("alice", _key(), response)] rag_routes._lesson_memory.clear() assert client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() == response assert pipeline.stream_calls == 1 def test_cache_key_includes_prompt_version(monkeypatch: pytest.MonkeyPatch) -> None: before = _key() monkeypatch.setattr(rag_routes, "LESSON_PROMPT_VERSION", "next") assert _key() != before def test_cache_skips_grounded_default_replacement(client: TestClient, pipeline: Pipeline) -> None: pipeline.verification = {"verified": False, "pCorrect": 0.2} first = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert first.status_code == 200 assert "Generated" not in str(first.json()["sections"]) assert pipeline.stream_calls == 2 assert pipeline.writes == [] def test_cache_skips_errors(client: TestClient, pipeline: Pipeline) -> None: pipeline.retrieval_error = RuntimeError("chroma down") assert client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).status_code == 503 pipeline.retrieval_error = None assert client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).status_code == 200 assert pipeline.stream_calls == 1 wait_for_writes(pipeline, 1) assert len(pipeline.writes) == 1 def test_cache_force_refresh_regenerates_only_own_lesson(client: TestClient, pipeline: Pipeline) -> None: alice_first = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() bob_first = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("bob")).json() alice_refreshed = client.post("/api/rag/lesson", json={**PAYLOAD, "forceRefresh": True}, headers=_as("alice")).json() alice_after = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() bob_after = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("bob")).json() assert pipeline.stream_calls == 3 wait_for_writes(pipeline, 3) assert alice_refreshed != alice_first assert alice_after == alice_refreshed == pipeline.store[("alice", _key())] assert bob_after == bob_first def test_cache_retrieval_is_shared_across_students(client: TestClient, pipeline: Pipeline) -> None: client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("bob")) assert pipeline.retrieval_calls == 1 assert pipeline.stream_calls == 2 # ─── dedupe ──────────────────────────────────────────────────────────────────── def _run_concurrently(*calls: Tuple[str, Dict[str, object], Dict[str, str]]) -> List[httpx.Response]: async def run() -> List[httpx.Response]: transport = httpx.ASGITransport(app=_app()) async with httpx.AsyncClient(transport=transport, base_url="http://test") as http: return await asyncio.gather(*(http.post(path, json=body, headers=headers) for path, body, headers in calls)) return asyncio.run(run()) def test_dedupe_concurrent_identical_requests_share_one_generation(pipeline: Pipeline) -> None: pipeline.stream_delay = 0.3 plain_a, plain_b, streamed = _run_concurrently( ("/api/rag/lesson", PAYLOAD, _as("alice")), ("/api/rag/lesson", {**PAYLOAD, "userId": "x"}, _as("alice")), ("/api/rag/lesson/stream", PAYLOAD, _as("alice")), ) assert pipeline.stream_calls == 1 assert plain_a.json() == plain_b.json() lesson_events = [data for event, data in parse_sse(streamed.text) if event == "lesson"] assert lesson_events == [plain_a.json()] def test_dedupe_is_per_student(pipeline: Pipeline) -> None: pipeline.stream_delay = 0.2 alice, bob = _run_concurrently(("/api/rag/lesson", PAYLOAD, _as("alice")), ("/api/rag/lesson", PAYLOAD, _as("bob"))) assert pipeline.stream_calls == 2 assert alice.json() != bob.json() def test_dedupe_generation_survives_stream_client_disconnect(pipeline: Pipeline) -> None: pipeline.stream_delay = 0.4 async def run() -> Tuple[httpx.Response, int]: transport = httpx.ASGITransport(app=_app()) async with httpx.AsyncClient(transport=transport, base_url="http://test") as http: abandoned = asyncio.create_task(http.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice"))) await asyncio.sleep(0.1) abandoned.cancel() await asyncio.gather(abandoned, return_exceptions=True) await asyncio.sleep(0.6) calls_before = pipeline.stream_calls return await http.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")), calls_before response, calls_before = asyncio.run(run()) assert calls_before == 1 assert response.status_code == 200 assert pipeline.stream_calls == 1 assert len(pipeline.writes) == 1 # ─── video ───────────────────────────────────────────────────────────────────── def test_video_lookup_runs_concurrently_with_generation(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_delay = 0.5 pipeline.video_delay = 0.5 started = time.perf_counter() response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) elapsed = time.perf_counter() - started assert response.status_code == 200 moments = dict(pipeline.events) assert moments["video_start"] < moments["generate_end"] assert elapsed < 0.9, f"video lookup was serialized after generation ({elapsed:.2f}s)" video = next(s for s in response.json()["sections"] if s["type"] == "video") assert video["videoId"] == "abc123" assert video["embedUrl"] == "https://www.youtube.com/embed/abc123" assert video["videos"] == VIDEOS def test_video_failure_still_yields_lesson(client: TestClient, pipeline: Pipeline) -> None: pipeline.video_error = RuntimeError("quota exceeded") response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 200 video = next(s for s in response.json()["sections"] if s["type"] == "video") assert video.get("videoId", "") == "" assert "videos" not in video assert len(response.json()["sections"]) == 7 def test_video_slow_lookup_is_capped(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setattr(rag_routes, "_LESSON_VIDEO_CAP_SEC", 0.2) pipeline.video_delay = 1.5 started = time.perf_counter() response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 200 assert time.perf_counter() - started < 1.0 assert next(s for s in response.json()["sections"] if s["type"] == "video").get("videoId", "") == "" # ─── stream ──────────────────────────────────────────────────────────────────── def test_stream_emits_stages_progress_and_lesson_matching_post(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_delay = 0.1 streamed = client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")) assert streamed.status_code == 200 assert streamed.headers["content-type"].startswith("text/event-stream") events = parse_sse(streamed.text) stages = [data["stage"] for event, data in events if event == "stage"] assert stages == ["retrieving", "generating", "verifying", "finalizing"] progress = [data for event, data in events if event == "progress"] assert {"phase": "thinking", "chars": 120} in progress assert {"phase": "writing", "chars": 480} in progress assert events[-1][0] == "lesson" plain = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() assert events[-1][1] == plain assert set(plain) >= { "sections", "retrievalConfidence", "retrievalBand", "retrievalMode", "needsReview", "sources", "activeModel", "jevVerification", "personalized", } SAMPLE_SSE_PATH.parent.mkdir(parents=True, exist_ok=True) SAMPLE_SSE_PATH.write_text(streamed.text, encoding="utf-8", newline="") def test_stream_memory_cache_hit_emits_cached_stage(client: TestClient, pipeline: Pipeline) -> None: plain = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() events = parse_sse(client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")).text) assert events == [("stage", {"stage": "cached"}), ("lesson", plain)] assert pipeline.stream_calls == 1 def test_stream_error_event_carries_status_and_detail(client: TestClient, pipeline: Pipeline) -> None: pipeline.chunks = [] plain = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) events = parse_sse(client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")).text) assert plain.status_code == 404 assert events[-1] == ("error", {"status": 404, "detail": plain.json()["detail"]}) assert [event for event, _ in events].count("lesson") == 0 def test_stream_inference_failure_maps_to_error_event(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: pipeline.stream_error = RuntimeError("reasoner down") def broken_client() -> SimpleNamespace: def generate_from_messages(_request: SimpleNamespace) -> str: raise rag_routes.InferenceConnectionError("chat down") return SimpleNamespace(generate_from_messages=generate_from_messages) monkeypatch.setattr(rag_routes, "_get_inference_client", broken_client) events = parse_sse(client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")).text) assert events[-1][0] == "error" assert events[-1][1]["status"] == 502 assert events[-1][1]["detail"]["error"] == "inference_connection_failed" def test_stream_validation_error_is_plain_http_status(client: TestClient) -> None: response = client.post("/api/rag/lesson/stream", json={**PAYLOAD, "quarter": "Q9"}, headers=_as("alice")) assert response.status_code == 422 assert response.headers["content-type"].startswith("application/json") def test_stream_pings_while_work_is_in_progress(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setattr(rag_routes, "_LESSON_PING_INTERVAL_SEC", 0.05) pipeline.stream_delay = 0.4 body = client.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")).text assert ": ping\n\n" in body assert parse_sse(body)[-1][0] == "lesson" # ─── fallback ────────────────────────────────────────────────────────────────── def test_fallback_to_chat_once_when_reasoner_stream_fails(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_error = RuntimeError("stream reset") response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 200 assert response.json()["activeModel"] == rag_routes.CHAT_MODEL assert _intro(response.json()).startswith("Fallback") assert len(pipeline.inference_requests) == 1 fallback = pipeline.inference_requests[0] assert fallback.model == rag_routes.CHAT_MODEL assert fallback.task_type == "rag_lesson" assert fallback.max_new_tokens == 4096 assert fallback.timeout_sec == 120 assert fallback.max_retries == 1 assert not {"temperature", "top_p"} & vars(fallback).keys() # A chat fallback is not persisted indefinitely: the next open retries the reasoner. assert pipeline.writes == [] def test_fallback_when_stream_exceeds_overall_cap(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setattr(rag_routes, "_LESSON_GENERATION_CAP_SEC", 0.2) pipeline.stream_delay = 1.0 started = time.perf_counter() response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 200 assert time.perf_counter() - started < 0.9 assert response.json()["activeModel"] == rag_routes.CHAT_MODEL assert len(pipeline.inference_requests) == 1 def _thinking(text: str) -> SimpleNamespace: return SimpleNamespace(type="content_block_delta", delta=SimpleNamespace(type="thinking_delta", thinking=text)) def _text(text: str) -> SimpleNamespace: return SimpleNamespace(type="content_block_delta", delta=SimpleNamespace(type="text_delta", text=text)) class FakeClaudeStream: """Stands in for open_stream(): a context manager that iterates raw events, then reports the final message.""" def __init__(self, events: List[SimpleNamespace], stop_reason: str = "end_turn") -> None: self.events = events self.stop_reason = stop_reason self.open_kwargs: Dict[str, object] = {} self.closed = False def open(self, messages: List[Dict[str, str]], **kwargs: object) -> "FakeClaudeStream": self.open_kwargs = {"messages": messages, **kwargs} return self def __enter__(self) -> "FakeClaudeStream": return self def __exit__(self, *_exc: object) -> None: self.closed = True def __iter__(self) -> Iterator[SimpleNamespace]: return iter(self.events) def get_final_message(self) -> SimpleNamespace: thinking = "".join(event.delta.thinking for event in self.events if event.delta.type == "thinking_delta") text = "".join(event.delta.text for event in self.events if event.delta.type == "text_delta") blocks = [SimpleNamespace(type="thinking", thinking=thinking)] if thinking else [] blocks += [SimpleNamespace(type="text", text=text)] if text else [] usage = SimpleNamespace(input_tokens=1200, output_tokens=len(thinking) + len(text)) return SimpleNamespace(stop_reason=self.stop_reason, stop_details=None, usage=usage, content=blocks) def test_fallback_contract_reasoner_stream_request_shape(monkeypatch: pytest.MonkeyPatch) -> None: fake = FakeClaudeStream([_thinking("think " * 10), _text('{"sections": '), _text("[]}")]) monkeypatch.setattr(rag_routes, "open_stream", fake.open) monkeypatch.setattr(rag_routes, "_LESSON_PROGRESS_INTERVAL_SEC", 0.0) progress: List[Tuple[str, int]] = [] text = rag_routes._stream_reasoner_lesson( [{"role": "user", "content": "x"}], REASONER, lambda phase, chars: progress.append((phase, chars)), threading.Event() ) assert text == '{"sections": []}' assert fake.open_kwargs["messages"] == [{"role": "user", "content": "x"}] assert fake.open_kwargs["model"] == REASONER assert fake.open_kwargs["max_tokens"] == rag_routes._REASONER_LESSON_MAX_TOKENS assert fake.open_kwargs["effort"] == rag_routes._REASONER_LESSON_EFFORT == "low" assert "temperature" not in fake.open_kwargs and "top_p" not in fake.open_kwargs assert fake.open_kwargs.get("max_retries", 0) == 0 timeout = fake.open_kwargs["timeout"] assert isinstance(timeout, Timeout) and timeout.read == 60.0 and timeout.connect == 10.0 assert progress[0] == ("thinking", 60) assert progress[-1] == ("writing", len('{"sections": []}')) assert fake.closed def test_fallback_triggered_by_truncated_or_empty_stream(monkeypatch: pytest.MonkeyPatch) -> None: truncated = FakeClaudeStream([_text('{"sections": [')], stop_reason="max_tokens") only_thinking = FakeClaudeStream([_thinking("only thinking")]) for fake in (truncated, only_thinking): monkeypatch.setattr(rag_routes, "open_stream", fake.open) with pytest.raises(RuntimeError): rag_routes._stream_reasoner_lesson([], REASONER, lambda phase, chars: None, threading.Event()) def test_fallback_triggered_by_refused_stream(monkeypatch: pytest.MonkeyPatch) -> None: refused = FakeClaudeStream([_text("I can't help with that.")], stop_reason="refusal") monkeypatch.setattr(rag_routes, "open_stream", refused.open) with pytest.raises(AIRefusalError): rag_routes._stream_reasoner_lesson([], REASONER, lambda phase, chars: None, threading.Event()) def test_stop_event_aborts_the_reasoner_stream_and_closes_it(monkeypatch: pytest.MonkeyPatch) -> None: fake = FakeClaudeStream([_text('{"sections": '), _text("[]}")]) monkeypatch.setattr(rag_routes, "open_stream", fake.open) stop = threading.Event() stop.set() with pytest.raises(TimeoutError): rag_routes._stream_reasoner_lesson([], REASONER, lambda phase, chars: None, stop) assert fake.closed # ─── validate / repair ───────────────────────────────────────────────────────── def test_validate_accepts_rich_lesson() -> None: assert rag_routes._lesson_deficits(_lesson_dict("Rich")) == [] def test_validate_reports_each_specific_deficit() -> None: thin = _lesson_dict("Thin") sections = {s["type"]: s for s in thin["sections"]} sections["introduction"]["content"] = "Short." sections["key_concepts"]["callouts"] = [] sections["worked_examples"]["examples"][0]["steps"] = ["Step 1: only one"] sections["important_notes"]["bulletPoints"] = ["one"] sections["try_it_yourself"]["practiceProblems"][0]["solution"] = "" sections["summary"]["content"] = "Short." deficits = " | ".join(rag_routes._lesson_deficits(thin)) for fragment in ("introduction", "callout", "worked_examples", "found 2", "important_notes", "try_it_yourself", "found 4", "summary"): assert fragment in deficits assert "key_concepts.content" not in deficits assert rag_routes._lesson_deficits({"explanation": "not json"}) == [ "The response was not valid lesson JSON with a 'sections' array." ] def test_repair_pass_fixes_thin_lesson_and_result_is_cached(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_outputs = [_lesson_json("Thin", rich=False)] response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() assert _intro(response).startswith("Repaired") assert response["activeModel"] == REASONER assert len(pipeline.inference_requests) == 1 repair = pipeline.inference_requests[0] assert repair.model == rag_routes.CHAT_MODEL assert repair.timeout_sec == 120 and repair.max_retries == 1 repair_prompt = repair.messages[-1]["content"] assert "[YOUR PREVIOUS OUTPUT]" in repair_prompt and "Thin introduction." in repair_prompt assert "worked_examples needs at least 3 examples" in repair_prompt wait_for_writes(pipeline, 1) assert len(pipeline.writes) == 1 def test_repair_runs_when_output_is_not_json(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_outputs = ["Sorry, here is a lesson in prose."] response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() assert _intro(response).startswith("Repaired") assert "not valid lesson JSON" in pipeline.inference_requests[0].messages[-1]["content"] def test_repair_failure_falls_back_to_grounded_defaults_uncached(client: TestClient, pipeline: Pipeline) -> None: pipeline.stream_outputs = [_lesson_json("Thin", rich=False)] pipeline.repair_output = _lesson_json("StillThin", rich=False) response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")).json() assert "Thin" not in str(response["sections"]) assert CHUNKS[0]["content"] in str(response["sections"]) assert len(response["sections"]) == 7 assert response["needsReview"] is True assert pipeline.writes == [] assert rag_routes._lesson_memory == {} # ─── profile ─────────────────────────────────────────────────────────────────── def test_profile_loader_fail_open_uses_generic_profile(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: prompts: List[str] = [] real_build = rag_routes.build_lesson_prompt def capture(**kwargs: object) -> str: prompts.append(str(kwargs["learner_profile"])) return real_build(**kwargs) def broken(uid: str, payload: object) -> str: raise RuntimeError("firestore unavailable") monkeypatch.setattr(rag_routes, "build_lesson_prompt", capture) pipeline.profile = broken response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 200 assert response.json()["personalized"] is False assert prompts == [rag_routes.GENERIC_LEARNER_PROFILE] def test_profile_loader_timeout_is_capped(client: TestClient, pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setattr(rag_routes, "_LEARNER_PROFILE_CAP_SEC", 0.1) def slow(uid: str, payload: object) -> str: time.sleep(1.0) return PROFILE pipeline.profile = slow started = time.perf_counter() response = client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) assert response.json()["personalized"] is False assert time.perf_counter() - started < 0.9 class _Doc: def __init__(self, fields: Optional[Dict[str, object]]) -> None: self.exists = fields is not None self._fields = fields def to_dict(self) -> Optional[Dict[str, object]]: return self._fields class _FakeDb: """collection/document chains resolve to a slash path looked up in `docs`.""" def __init__(self, docs: Dict[str, Dict[str, object]], path: str = "") -> None: self.docs = docs self.path = path def collection(self, name: str) -> "_FakeDb": return _FakeDb(self.docs, f"{self.path}/{name}".lstrip("/")) document = collection def get(self) -> _Doc: return _Doc(self.docs.get(self.path)) def test_profile_loader_builds_compact_profile_without_pii(monkeypatch: pytest.MonkeyPatch) -> None: docs = { "users/alice": { "name": "Alice Santos", "email": "alice@example.com", "grade": "Grade 11", "track": "STEM", "iarAssessmentState": "placed", "startingQuarterG11": "Q1", "iarTopicClassifications": {"Functions": "NeedsReview"}, "topicScores": {"Functions": 45}, "assessmentResults": {"weakTopics": ["Domain and range"], "strongTopics": ["Logic"]}, "recommendedPace": "support_intensive", "overallRisk": "High", "atRiskSubjects": ["General Mathematics"], }, "studentProgress/alice/topics/gm-q1-rational": {"mastery_level": "developing", "best_score": 60}, } monkeypatch.setattr(rag_routes, "_firestore_db", lambda: _FakeDb(docs)) import services.memory_service as memory_service tutor = SimpleNamespace(weak_topics=["asymptotes"], recurring_mistakes=["divides by zero"], explanation_depth="basic") monkeypatch.setattr(memory_service, "load_profile", lambda uid: tutor) profile = rag_routes._load_learner_profile("alice", rag_routes.RagLessonRequest(**PAYLOAD)) for fragment in ("STEM", "placed", "Functions=NeedsReview", "Domain and range", "support_intensive", "at risk in General Mathematics", "mastery developing", "divides by zero", "basic"): assert fragment in profile assert "Alice" not in profile and "alice@" not in profile assert len(profile) <= rag_routes._LEARNER_PROFILE_MAX_CHARS def test_profile_prompt_contains_learner_profile_section() -> None: prompt = rag_routes.build_lesson_prompt( lesson_title="Rational Functions", competency="M11GM-Ib-1", grade_level="Grade 11", subject="General Mathematics", quarter=1, learner_level=None, module_unit=None, curriculum_chunks=CHUNKS, learner_profile=PROFILE, ) assert "[LEARNER PROFILE]\n" + PROFILE in prompt assert "pesos" in prompt and "$...$" in prompt and "7 sections" in prompt # ─── prefetch ────────────────────────────────────────────────────────────────── def test_prefetch_returns_cached_without_generating(client: TestClient, pipeline: Pipeline) -> None: pipeline.store[("alice", _key())] = _lesson_dict("Stored") response = client.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) assert response.status_code == 202 assert response.json() == {"status": "cached"} assert pipeline.stream_calls == 0 def test_prefetch_anonymous_is_skipped(client: TestClient, pipeline: Pipeline) -> None: assert client.post("/api/rag/lesson/prefetch", json=PAYLOAD).json() == {"status": "skipped"} assert pipeline.stream_calls == 0 def test_prefetch_queued_generates_once_and_stream_joins_in_flight(pipeline: Pipeline) -> None: pipeline.stream_delay = 0.4 async def run() -> Tuple[httpx.Response, httpx.Response, httpx.Response, httpx.Response]: transport = httpx.ASGITransport(app=_app()) async with httpx.AsyncClient(transport=transport, base_url="http://test") as http: queued = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) again = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) await asyncio.sleep(0.1) streamed = await http.post("/api/rag/lesson/stream", json=PAYLOAD, headers=_as("alice")) await asyncio.sleep(0.05) after = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) return queued, again, streamed, after queued, again, streamed, after = asyncio.run(run()) assert queued.status_code == 202 and queued.json() == {"status": "queued"} assert again.json() == {"status": "in_progress"} events = parse_sse(streamed.text) assert events[0] == ("stage", {"stage": "generating"}) assert events[-1][0] == "lesson" assert after.json() == {"status": "cached"} assert pipeline.stream_calls == 1 assert len(pipeline.writes) == 1 assert rag_routes._prefetch_uids == set() def test_prefetch_per_student_limit_returns_skipped(pipeline: Pipeline) -> None: pipeline.stream_delay = 0.3 other_lesson = {**PAYLOAD, "lessonTitle": "Graphs of Rational Functions", "lessonId": "gm-q1-graphs"} async def run() -> Tuple[httpx.Response, httpx.Response, httpx.Response]: transport = httpx.ASGITransport(app=_app()) async with httpx.AsyncClient(transport=transport, base_url="http://test") as http: first = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) second = await http.post("/api/rag/lesson/prefetch", json=other_lesson, headers=_as("alice")) other_student = await http.post("/api/rag/lesson/prefetch", json=other_lesson, headers=_as("bob")) await asyncio.sleep(0.6) return first, second, other_student first, second, other_student = asyncio.run(run()) assert first.json() == {"status": "queued"} assert second.json() == {"status": "skipped"} assert other_student.json() == {"status": "queued"} assert pipeline.stream_calls == 2 def test_prefetch_global_limit_returns_skipped(pipeline: Pipeline, monkeypatch: pytest.MonkeyPatch) -> None: monkeypatch.setattr(rag_routes, "_prefetch_semaphore", asyncio.Semaphore(1)) pipeline.stream_delay = 0.3 async def run() -> Tuple[httpx.Response, httpx.Response, httpx.Response]: transport = httpx.ASGITransport(app=_app()) async with httpx.AsyncClient(transport=transport, base_url="http://test") as http: first = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("alice")) await asyncio.sleep(0.05) second = await http.post("/api/rag/lesson/prefetch", json=PAYLOAD, headers=_as("bob")) interactive = await http.post("/api/rag/lesson", json=PAYLOAD, headers=_as("carol")) await asyncio.sleep(0.4) return first, second, interactive first, second, interactive = asyncio.run(run()) assert first.json() == {"status": "queued"} assert second.json() == {"status": "skipped"} assert interactive.status_code == 200 # ─── timings ─────────────────────────────────────────────────────────────────── def test_lesson_logs_stage_timings_once_per_lesson(client: TestClient, pipeline: Pipeline, caplog: pytest.LogCaptureFixture) -> None: with caplog.at_level("INFO", logger="mathpulse.rag"): client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) client.post("/api/rag/lesson", json=PAYLOAD, headers=_as("alice")) lines = [r.getMessage() for r in caplog.records if r.getMessage().startswith("rag_lesson timings")] assert len(lines) == 2 assert "cache=miss" in lines[0] for stage in ("retrieve_ms=", "generate_ms=", "verify_ms=", "video_ms=", "total_ms="): assert stage in lines[0] assert "cache=hit" in lines[1] print(lines[0])