test: 补充知识处理、检索与 Agent 回归用例

覆盖切块、向量检索、工具调用和历史预算；模型相关用例使用桩。

backend/tests/utils/test_offtopic_exercise.py

@@ -0,0 +1,125 @@
+"""
+Copyright 2024-2026 ReplyPier
+
+Licensed under the Apache License, Version 2.0 (the "License");
+you may not use this file except in compliance with the License.
+You may obtain a copy of the License at
+
+    http://www.apache.org/licenses/LICENSE-2.0
+
+Unless required by applicable law or agreed to in writing, software
+distributed under the License is distributed on an "AS IS" BASIS,
+WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+See the License for the specific language governing permissions and
+limitations under the License.
+
+The off-topic exercise detector exists because prompt text demonstrably failed:
+a long algorithms brief was answered in full in production under three
+successive policy wordings. It is deliberately narrow, so these tests weigh
+PRECISION far above recall — a wrongly blocked customer is worse than a missed
+exercise, which the prompt policy still catches on the way through.
+"""
+
+import pytest
+
+from app.utils.guardrails import detect_offtopic_exercise
+
+# The exact brief that got through in production on 2026-07-31.
+PROD_BRIEF = """Problem: Median in a Stream with Deletions
+Design a data structure that supports the following operations, each in O(log n)
+time: insert(x) — add value x to the collection; delete(x) — remove one
+occurrence of value x from the collection (guaranteed to exist); find_median() —
+return the median of all currently present values in O(1) time.
+Constraints: Values are not necessarily distinct (duplicates allowed). n can be
+up to 10^6, and operations are interleaved arbitrarily.
+The naive two-heap median trick handles insert and find_median well, but breaks
+down for delete.
+Your task: Design a structure that supports all three operations within the time
+bounds. State and justify the time and space complexity of each operation.
+Bonus: how would your design change if you also needed find_kth(k)?"""
+
+# Long, technical, and unmistakably about a real product.
+LONG_INTEGRATION_QUESTION = """We're rolling your widget out across three
+storefronts and I want to get the architecture right before we commit.
+Each store has its own subdomain and its own knowledge base, and we'd like a
+single dashboard view across all of them. Right now we install the widget
+script per store, but the api key seems to be scoped per organisation rather
+than per site, so I can't tell the conversations apart in the inbox.
+Is there a recommended setup for multi-brand deployments? Should we create one
+organisation per storefront and accept separate billing, or is there a way to
+segment within one account? We also need the webhook payloads to identify the
+originating store so our CRM can route them.""" + " padding. " * 20
+
+LONG_TRACEBACK = """The worker keeps dying in production and I can't work out why.
+Here is what we get, repeated every few minutes:
+Traceback (most recent call last):
+  File "app/worker.py", line 42, in <module>
+    client.connect()
+  File "app/net.py", line 118, in connect
+    raise ConnectionRefusedError(errno.ECONNREFUSED, os.strerror(errno.ECONNREFUSED))
+ConnectionRefusedError: [Errno 111] Connection refused
+This started after we moved the container to a new host. The docker compose
+setup is unchanged and redis is definitely running. Any idea what we should
+check first? It fails on every restart.""" + " More detail follows. " * 12
+
+
+class TestCatchesTheAbuse:
+    def test_the_production_brief(self):
+        assert detect_offtopic_exercise(PROD_BRIEF) is True
+
+    def test_still_caught_when_org_name_is_known(self):
+        assert detect_offtopic_exercise(PROD_BRIEF, ["ReplyPier"]) is True
+
+
+class TestPrecision:
+    """Everything here is legitimate traffic and must never be caught."""
+
+    def test_long_product_question_with_business_anchors(self):
+        assert detect_offtopic_exercise(LONG_INTEGRATION_QUESTION) is False
+
+    def test_long_pasted_traceback(self):
+        assert detect_offtopic_exercise(LONG_TRACEBACK) is False
+
+    def test_exercise_shaped_but_about_the_product(self):
+        """A genuine scaling question that happens to use the same vocabulary."""
+        text = (
+            "We push about 10^6 conversations a month through your webhook and "
+            "the ordering matters to us. What data structure do you use "
+            "internally for the queue, and what is the time complexity of a "
+            "replay? We need worst-case guarantees before we design around it. "
+            "Constraints: we cannot lose events, and duplicates allowed is not "
+            "acceptable for our billing reconciliation."
+        ) + " Extra context. " * 20
+        assert detect_offtopic_exercise(text) is False
+
+    def test_short_exercise_is_left_to_the_prompt(self):
+        """Below the length bar the policy block handles it — the detector is
+        for the long briefs the prompt cannot hold."""
+        short = (
+            "Design a data structure with insert and getRandom in O(1). "
+            "Implement the class and give the time complexity."
+        )
+        assert detect_offtopic_exercise(short) is False
+
+    @pytest.mark.parametrize("text", [
+        "hi",
+        "",
+        "How much does the Pro plan cost per seat, and is there a free trial?",
+        "sudo docker compose up -d fails with permission denied, any ideas?",
+    ])
+    def test_ordinary_messages(self, text):
+        assert detect_offtopic_exercise(text) is False
+
+    def test_none_is_safe(self):
+        assert detect_offtopic_exercise(None) is False
+
+
+class TestThresholds:
+    def test_needs_enough_markers(self):
+        """Long and technical is not enough on its own."""
+        text = ("We use an algorithm to sort the results. " + "Filler text here. " * 40)
+        assert detect_offtopic_exercise(text) is False
+
+    def test_org_name_mention_anchors_it(self):
+        text = PROD_BRIEF + "\nWe need this for our ReplyPier integration."
+        assert detect_offtopic_exercise(text, ["ReplyPier"]) is False