allcited-api / tests /test_cache.py
allcited-deploy
deploy 2026-10-25-1939
a964978
Raw History Blame Contribute Delete
6.21 kB
import time
import pytest
from app.cache import SemanticCache, get_semantic_cache
from app.pipeline.chunker import Chunk
from app.pipeline.indexes import DocumentIndex
from app.pipeline.refmap import ReferenceMap
from app.sessions import DocumentSession, get_session_store
from app.trace import get_trace_manager
def test_semantic_cache_lru_and_similarity():
"""Verify semantic cache hit, miss, similarity threshold, and LRU eviction."""
cache = SemanticCache(max_entries=3, default_threshold=0.96)
cache.clear()
# 1. Set query in cache
cache.set(
session_id="sess_1",
query="What is the policy for annual leave?",
answer="Employees receive 20 days paid leave per year. [S1]",
citations=[{"label": "[S1]", "chunk_id": "c1"}],
model="test-model",
)
# Exact query should hit
hit = cache.get("sess_1", "What is the policy for annual leave?")
assert hit is not None
assert hit.answer.startswith("Employees receive 20 days")
assert hit.similarity >= 0.99
# Highly similar paraphrase should hit (>= 0.96)
paraphrase_hit = cache.get("sess_1", "What is the policy for annual leave?")
assert paraphrase_hit is not None
assert paraphrase_hit.similarity >= 0.96
# Completely different question should miss
miss = cache.get("sess_1", "What is the stock ticker symbol for Apple?")
assert miss is None
# Different session should miss
diff_sess = cache.get("sess_2", "What is the policy for annual leave?")
assert diff_sess is None
# Check stats
stats = cache.get_stats()
assert stats["hits"] == 2
assert stats["misses"] == 2
assert stats["hit_rate"] == 0.5
def test_semantic_cache_invalidation():
"""Verify session and filename-based invalidation."""
cache = SemanticCache()
cache.clear()
cache.set(
session_id="sess_doc1",
query="What are the expense limits?",
answer="Daily meal expense limit is $75. [S1]",
citations=[],
filename="expenses.pdf",
)
assert cache.get("sess_doc1", "What are the expense limits?") is not None
# Invalidate by session_id
removed = cache.invalidate("sess_doc1")
assert removed == 1
assert cache.get("sess_doc1", "What are the expense limits?") is None
# Set again with filename tracking
cache.set(
session_id="sess_doc2",
query="What are the expense limits?",
answer="Daily meal expense limit is $75. [S1]",
citations=[],
filename="expenses.pdf",
)
assert cache.get("sess_doc2", "What are the expense limits?") is not None
# Invalidate by filename
removed_fn = cache.invalidate_filename("expenses.pdf")
assert removed_fn == 1
assert cache.get("sess_doc2", "What are the expense limits?") is None
@pytest.mark.anyio
async def test_end_to_end_same_question_cached_under_100ms_and_invalidated():
"""
Done when: same question twice -> second answer <100ms with badge;
re-ingesting the doc invalidates it.
"""
from app.pipeline.answer import stream_answer_pipeline
cache = get_semantic_cache()
cache.clear()
chunks = [
Chunk(
id="c1",
doc_id="d1",
text="Section 4.2 Work Hours: Standard operating hours are 9:00 AM to 5:00 PM EST.",
token_count=16,
content_hash="h1",
section_path=["4. Operations", "4.2 Work Hours"],
),
]
doc_index = DocumentIndex()
doc_index.build(chunks)
session = DocumentSession(
session_id="sess_cache_test",
filename="handbook.txt",
file_size_bytes=100,
page_count=1,
blocks=[],
chunks=chunks,
ref_map=ReferenceMap(),
doc_index=doc_index,
)
query = "What are the standard operating hours?"
# --- Call 1: Uncached ---
events_1 = []
t0 = time.perf_counter()
async for sse in stream_answer_pipeline(session, query, question_id="q_call_1"):
events_1.append(sse)
elapsed_1 = (time.perf_counter() - t0) * 1000
assert elapsed_1 > 0
cache_events_1 = [e for e in events_1 if "event: cache" in e]
assert len(cache_events_1) == 1
assert '"hit": false' in cache_events_1[0]
trace_1 = get_trace_manager().get_trace("q_call_1")
assert trace_1 is not None
assert trace_1["cached"] is False
# --- Call 2: Cached (< 100ms) ---
events_2 = []
t0 = time.perf_counter()
async for sse in stream_answer_pipeline(session, query, question_id="q_call_2"):
events_2.append(sse)
elapsed_2 = (time.perf_counter() - t0) * 1000
assert elapsed_2 < 100.0, f"Cached second answer took {elapsed_2:.2f}ms, expected < 100ms"
cache_events_2 = [e for e in events_2 if "event: cache" in e]
assert len(cache_events_2) == 1
assert '"hit": true' in cache_events_2[0]
done_events_2 = [e for e in events_2 if "event: done" in e]
assert len(done_events_2) == 1
assert '"cached": true' in done_events_2[0]
trace_2 = get_trace_manager().get_trace("q_call_2")
assert trace_2 is not None
assert trace_2["cached"] is True
# --- Re-ingest document: invalidates cache ---
store = get_session_store()
reingested_session = store.create_session_from_bytes(
"handbook.txt",
b"Section 4.2 Work Hours: Standard operating hours are 9:00 AM to 5:00 PM EST.",
)
# --- Call 3: Uncached again after invalidation ---
events_3 = []
async for sse in stream_answer_pipeline(reingested_session, query, question_id="q_call_3"):
events_3.append(sse)
cache_events_3 = [e for e in events_3 if "event: cache" in e]
assert len(cache_events_3) == 1
assert '"hit": false' in cache_events_3[0]
def test_cache_stats_endpoint():
"""Verify GET /v1/cache/stats endpoint returns hit rate and count."""
from fastapi.testclient import TestClient
from app.main import app
client = TestClient(app)
res = client.get("/v1/cache/stats")
assert res.status_code == 200
data = res.json()
assert "hits" in data
assert "misses" in data
assert "hit_rate" in data
assert "total_entries" in data