fix(rag): apply channel scope before retrieval limit

backend/app/rag.py

@@ -93,7 +93,8 @@ def _store_conversation_turn(self, conversation_id: str, user_id: str, channel_i
 
     def _agent_payload(self, question: str, channel_id=None, conversation_id=None,
                        request_id=None, slack=None):
-        results = vector_store.similarity_search(question)
+        search_options = {"channel_id": channel_id} if channel_id else {}
+        results = vector_store.similarity_search(question, **search_options)
         if channel_id:
             results = [r for r in results if r.get("metadata", {}).get("channel_id") == channel_id]
         identifier = request_id or agent_runtime.request_id()

backend/app/vector_store.py

@@ -67,14 +67,16 @@ def add_documents(self, texts: List[str], metadatas: List[Dict[str, Any]] = None
 
         return self.vector_store.add_texts(texts=texts, metadatas=metadatas)
 
-    def similarity_search(self, query: str, k: int = None) -> List[Dict[str, Any]]:
-        """Search for similar documents"""
+    def similarity_search(self, query: str, k: int = None, *,
+                          channel_id: str = None) -> List[Dict[str, Any]]:
+        """Search within a channel before limiting results, or globally if unscoped."""
         self._ensure_initialized()
         if not self.vector_store:
             raise RuntimeError("Vector store not initialized")
 
         k = k or settings.RETRIEVAL_TOP_K
-        results = self.vector_store.similarity_search(query, k=k)
+        search_options = {"filter": {"channel_id": channel_id}} if channel_id else {}
+        results = self.vector_store.similarity_search(query, k=k, **search_options)
 
         return [
             {

backend/tests/test_agent_runtime.py

@@ -39,6 +39,7 @@ def test_rag_routes_retrieved_context_and_stable_slack_identity(monkeypatch):
     submit = Mock(return_value={"id": "job-1", "status": "pending"})
     monkeypatch.setattr(agent_runtime, "submit", submit)
     module.rag_engine.submit_slack("Question", "C1", "team:thread:C1:1.2", "event-1", "1.2")
+    vector.similarity_search.assert_called_once_with("Question", channel_id="C1")
     payload = submit.call_args.args[0]
     assert payload["request_id"] == "event-1"
     assert payload["conversation_id"] == "team:thread:C1:1.2"

backend/tests/test_retrieval.py

@@ -0,0 +1,110 @@
+"""Channel retrieval regressions using local Chroma and fixed embeddings."""
+import importlib.util
+from pathlib import Path
+import sys
+from types import SimpleNamespace
+from unittest.mock import Mock
+
+import chromadb
+from chromadb.config import Settings as ChromaSettings
+from langchain_community.vectorstores import Chroma
+from langchain_core.embeddings import Embeddings
+import pytest
+from chromadb.telemetry.product.posthog import Posthog
+
+from app.config import settings
+from app.rules import RuleEngine
+from app.vector_store import VectorStore
+
+
+class FixtureEmbeddings(Embeddings):
+    def embed_documents(self, texts):
+        distances = {
+            "Other channel first": 0.0,
+            "Other channel second": 0.1,
+            "Shared document": 0.2,
+            "Target channel first": 1.0,
+            "Target channel second": 2.0,
+        }
+        return [[distances[text], 1.0] for text in texts]
+
+    def embed_query(self, text):
+        return [0.0, 1.0]
+
+
+@pytest.fixture
+def vector(tmp_path, monkeypatch):
+    monkeypatch.setattr(settings, "RETRIEVAL_TOP_K", 2)
+    monkeypatch.setattr(Posthog, "capture", lambda *args, **kwargs: None)
+    client = chromadb.PersistentClient(
+        path=str(tmp_path), settings=ChromaSettings(anonymized_telemetry=False),
+    )
+    store = VectorStore()
+    store.vector_store = Chroma(
+        client=client, collection_name="retrieval-fixture",
+        embedding_function=FixtureEmbeddings(),
+    )
+    store._initialized = True
+    store.vector_store.add_texts(
+        ["Other channel first", "Other channel second", "Shared document",
+         "Target channel first", "Target channel second"],
+        [{"source": "slack", "channel_id": "COTHER"},
+         {"source": "slack", "channel_id": "COTHER"},
+         {"source": "api"},
+         {"source": "slack", "channel_id": "CTARGET"},
+         {"source": "slack", "channel_id": "CTARGET"}],
+    )
+    return store
+
+
+def test_channel_filter_runs_before_top_k(vector):
+    results = vector.similarity_search("How do I connect?", channel_id="CTARGET")
+    assert [item["content"] for item in results] == [
+        "Target channel first", "Target channel second",
+    ]
+
+
+def test_unscoped_search_preserves_global_ranking_and_explicit_k(vector):
+    results = vector.similarity_search("How do I connect?", k=3)
+    assert [item["content"] for item in results] == [
+        "Other channel first", "Other channel second", "Shared document",
+    ]
+
+
+def test_empty_channel_does_not_fall_back_to_global_results(vector):
+    assert vector.similarity_search("How do I connect?", channel_id="CEMPTY") == []
+
+
+@pytest.mark.parametrize("path", ["api", "slack"])
+def test_rag_payload_uses_scoped_matches(vector, monkeypatch, path):
+    monkeypatch.setitem(sys.modules, "app.vector_store", SimpleNamespace(vector_store=vector))
+    spec = importlib.util.spec_from_file_location(
+        "retrieval_rag_under_test", Path(__file__).parents[1] / "app/rag.py",
+    )
+    module = importlib.util.module_from_spec(spec)
+    spec.loader.exec_module(module)
+    monkeypatch.setattr(module.rules, "rule_engine", RuleEngine([]))
+    generate = Mock(return_value={"answer": "fixture answer"})
+    submit = Mock(return_value={"id": "fixture-job", "status": "pending"})
+    monkeypatch.setattr(module.agent_runtime, "generate", generate)
+    monkeypatch.setattr(module.agent_runtime, "submit", submit)
+
+    if path == "api":
+        module.rag_engine.generate_response(
+            "How do I connect?", channel_id="CTARGET", request_id="fixture-api",
+        )
+        payload = generate.call_args.args[0]
+    else:
+        module.rag_engine.submit_slack(
+            "How do I connect?", "CTARGET", "fixture:thread", "fixture-slack", "1.2",
+        )
+        payload = submit.call_args.args[0]
+        assert payload["slack"] == {"channel": "CTARGET", "thread_ts": "1.2"}
+
+    assert payload["context"] == (
+        "From slack:\nTarget channel first\n\nFrom slack:\nTarget channel second"
+    )
+    assert payload["sources"] == [
+        {"source": "slack", "channel_id": "CTARGET"},
+        {"source": "slack", "channel_id": "CTARGET"},
+    ]

backend/tests/test_rules.py

@@ -195,7 +195,7 @@ def test_unmatched_slack_question_uses_retrieval_without_pinned_answer(self, mon
         self._install_engine(monkeypatch, engine(rule("vpn", ["reset vpn"])))
         monkeypatch.setattr(
             rag_module.vector_store, "similarity_search",
-            lambda _: [{"content": "Doc", "metadata": {"source": "slack", "channel_id": "C1"}}],
+            lambda _, **kwargs: [{"content": "Doc", "metadata": {"source": "slack", "channel_id": "C1"}}],
         )
         submitted = {}
         def fake_submit(payload):

mise.toml

@@ -51,7 +51,7 @@ run = "docker compose ps"
 [tasks.test-backend]
 description = "Run backend tests"
 dir = "backend"
-run = "uv run pytest tests/test_file_extraction_units.py tests/test_logging_config.py tests/test_logging.py tests/test_agent_runtime.py tests/test_rules.py tests/test_conversations.py tests/test_documents.py -v"
+run = "uv run pytest tests/test_file_extraction_units.py tests/test_logging_config.py tests/test_logging.py tests/test_agent_runtime.py tests/test_rules.py tests/test_conversations.py tests/test_documents.py tests/test_retrieval.py -v"
 
 [tasks.migrate]
 description = "Run database migrations (upgrade to latest)"