fix: close semantic isolation review gaps

This commit is contained in:
2026-08-08 23:53:37 +02:00
parent 2c4534d968
commit fa151c5ea5
5 changed files with 158 additions and 4 deletions
+80
View File
@@ -289,11 +289,30 @@ def test_search_filters_by_workspace_and_allowed_record_kinds():
assert query_call[2]["filter"] == {
"must": [
{"key": "workspace_id", "match": {"value": "demo"}},
{"key": "kind", "match": {"any": ["memory"]}},
{"key": "record_kind", "match": {"any": ["memory"]}},
]
}
def test_search_excludes_inconsistent_semantic_kind_in_bound_workspace():
fake = FakeQdrantHttp()
store = _store(fake)
store.upsert("memory", [_write_record("memory:1", "memory")])
contaminated = next(iter(fake.points.values())).copy()
contaminated["id"] = point_id("demo", "evidence", "memory:contaminated")
contaminated["payload"] = {
**contaminated["payload"],
"kind": "evidence",
"record_key": "memory:contaminated",
}
fake.points[contaminated["id"]] = contaminated
hits = store.search(["memory"], [0.2] * 1024, limit=5, kinds=["memory"])
assert [hit.id for hit in hits] == ["memory:1"]
def test_existing_hashes_health_and_exact_generation_inventory_and_delete():
fake = FakeQdrantHttp()
store = _store(fake)
@@ -328,6 +347,46 @@ def test_existing_hashes_health_and_exact_generation_inventory_and_delete():
assert health.dimension_compatible is True
def test_evidence_inventory_and_delete_ignore_inconsistent_semantic_kind():
fake = FakeQdrantHttp()
store = _store(fake)
generation = "gen:" + "1" * 32
contaminated_generation = "gen:" + "2" * 32
store.upsert("evidence", [
_write_record(
f"demo:{generation}:chunk:1",
"evidence",
metadata={
"workspace_id": "demo",
"vector_generation": generation,
"document_id": "doc:1",
},
),
])
contaminated_delete = next(iter(fake.points.values())).copy()
contaminated_delete["id"] = point_id("demo", "memory", "evidence:contaminated-delete")
contaminated_delete["payload"] = {
**contaminated_delete["payload"],
"kind": "memory",
"record_key": "evidence:contaminated-delete",
}
fake.points[contaminated_delete["id"]] = contaminated_delete
contaminated_list = next(iter(fake.points.values())).copy()
contaminated_list["id"] = point_id("demo", "memory", "evidence:contaminated-list")
contaminated_list["payload"] = {
**contaminated_list["payload"],
"kind": "memory",
"record_key": "evidence:contaminated-list",
"vector_generation": contaminated_generation,
}
fake.points[contaminated_list["id"]] = contaminated_list
assert store.list_evidence_generations("evidence", "demo") == [generation]
assert store.delete_generation("evidence", generation, "demo") == 1
assert contaminated_delete["id"] in fake.points
assert contaminated_list["id"] in fake.points
def test_metadata_search_rejects_a_workspace_id_different_from_the_bound_adapter():
fake = FakeQdrantHttp()
store = _store(fake)
@@ -444,6 +503,7 @@ def test_delete_kinds_is_workspace_scoped_and_preserves_other_semantic_kinds():
assert delete_call[2]["filter"] == {
"must": [
{"key": "workspace_id", "match": {"value": "demo"}},
{"key": "kind", "match": {"any": ["memory"]}},
{"key": "record_kind", "match": {"any": ["memory"]}},
]
}
@@ -457,6 +517,26 @@ def test_delete_kinds_is_workspace_scoped_and_preserves_other_semantic_kinds():
}
def test_existing_hashes_and_delete_kinds_ignore_inconsistent_semantic_kind():
fake = FakeQdrantHttp()
store = _store(fake)
store.upsert("memory", [_write_record("memory:1", "memory")])
contaminated = next(iter(fake.points.values())).copy()
contaminated["id"] = point_id("demo", "evidence", "memory:contaminated")
contaminated["payload"] = {
**contaminated["payload"],
"kind": "evidence",
"record_key": "memory:contaminated",
}
fake.points[contaminated["id"]] = contaminated
assert store.existing_hashes("memory", ["memory"]) == {
"memory:1": "sha256:" + "a" * 64,
}
assert store.delete_kinds("memory", ["memory"]) == 1
assert contaminated["id"] in fake.points
def test_sanitizes_timeout_and_malformed_responses():
fake = FakeQdrantHttp()
store = _store(fake)
+9
View File
@@ -125,6 +125,7 @@ class QdrantVectorStore:
if not allowed_record_kinds:
return []
filter_must = self._workspace_filter()
filter_must.append(self._semantic_kind_filter(allowed_record_kinds))
filter_must.append({"key": "record_kind", "match": {"any": allowed_record_kinds}})
if metadata_filter is not None:
if set(metadata_filter) != {"vector_generation", "document_ids", "workspace_id"}:
@@ -166,6 +167,7 @@ class QdrantVectorStore:
points = self._scroll(
[
*self._workspace_filter(),
self._semantic_kind_filter(kinds),
{"key": "record_kind", "match": {"any": sorted(kinds)}},
]
)
@@ -213,6 +215,7 @@ class QdrantVectorStore:
validate_collection_kinds(collection, kinds)
must = [
*self._workspace_filter(),
self._semantic_kind_filter(kinds),
{"key": "record_kind", "match": {"any": sorted(kinds)}},
]
before = len(self._scroll(must))
@@ -231,6 +234,7 @@ class QdrantVectorStore:
self._require_bound_workspace(workspace_id)
must = [
*self._workspace_filter(),
{"key": "kind", "match": {"value": "evidence"}},
{"key": "record_kind", "match": {"any": ["evidence"]}},
{"key": "vector_generation", "match": {"value": generation}},
]
@@ -253,6 +257,7 @@ class QdrantVectorStore:
points = self._scroll(
[
*self._workspace_filter(),
{"key": "kind", "match": {"value": "evidence"}},
{"key": "record_kind", "match": {"any": ["evidence"]}},
]
)
@@ -268,6 +273,10 @@ class QdrantVectorStore:
def _workspace_filter(self) -> list[dict]:
return [{"key": "workspace_id", "match": {"value": self._workspace_id}}]
def _semantic_kind_filter(self, record_kinds: list[str]) -> dict:
semantic_kinds = sorted({qdrant_semantic_kind(kind) for kind in record_kinds})
return {"key": "kind", "match": {"any": semantic_kinds}}
def _require_bound_workspace(self, workspace_id: str) -> None:
if workspace_id != self._workspace_id:
raise VectorStoreError("Evidence workspace namespace does not match bound workspace")
+67 -2
View File
@@ -150,12 +150,28 @@ fi
if [ "$1" = compose ] && [ "$2" = --project-name ]; then
case "$4" in
ps)
if [ "${QDRANT_RUNNING:-1}" = 1 ]; then
qdrant_state=${QDRANT_RUNNING:-1}
if [ -n "${QDRANT_STATE_FILE:-}" ] && [ -f "$QDRANT_STATE_FILE" ]; then
qdrant_state=$(cat "$QDRANT_STATE_FILE")
fi
if [ "$qdrant_state" = 1 ]; then
printf '%s\n' qdrant-container
fi
exit 0
;;
stop|start)
stop)
if [ -n "${QDRANT_STATE_FILE:-}" ]; then
printf '%s\n' 0 >"$QDRANT_STATE_FILE"
fi
if [ "${STOP_MARKS_STOPPED_THEN_FAIL:-0}" = 1 ]; then
exit 42
fi
exit 0
;;
start)
if [ -n "${QDRANT_STATE_FILE:-}" ]; then
printf '%s\n' 1 >"$QDRANT_STATE_FILE"
fi
exit 0
;;
esac
@@ -525,4 +541,53 @@ grep -q "compose --project-name $project stop qdrant" "$rollback_log"
grep -q "compose --project-name $project start qdrant" "$rollback_log"
test ! -e "$tmp/locks/$lock_name"
assert_ambiguous_stop_reconciled() {
operation=$1
operation_log="$tmp/$operation-ambiguous-stop.log"
operation_state="$tmp/$operation-qdrant-state"
foreign_lock="$tmp/locks/foreign-$operation-lock"
printf '%s\n' 1 >"$operation_state"
mkdir -p "$foreign_lock"
printf '%s\n' foreign-owner >"$foreign_lock/owner"
printf '%s\n' foreign-volume >"$foreign_lock/volume"
if [ "$operation" = backup ]; then
set -- ./scripts/vector-backup.sh --project-name "$project" \
--output "$tmp/ambiguous-stop-backup.tar"
else
set -- ./scripts/vector-restore.sh --project-name "$project" \
--input "$restore_input" --confirm-project "$project"
fi
if PATH="$fakebin:$PATH" DOCKER_LOG="$operation_log" PROJECT_NAME="$project" \
VOLUME_ROOT="$volume_root" QDRANT_STATE_FILE="$operation_state" \
STOP_MARKS_STOPPED_THEN_FAIL=1 \
HELPER_IMAGE="qdrant/qdrant:v1.18.2@sha256:75eab8c4ba42096724fdcfde8b4de0b5713d529dde32f285a1f86fdcb2c9e50c" \
"$@" >"$tmp/$operation-ambiguous-stop.out" \
2>"$tmp/$operation-ambiguous-stop.err"; then
echo "$operation hid an ambiguous qdrant stop failure" >&2
exit 1
fi
grep -q "compose --project-name $project stop qdrant" "$operation_log"
grep -q "compose --project-name $project start qdrant" "$operation_log"
test "$(cat "$operation_state")" = 1
test ! -e "$tmp/locks/$lock_name"
test "$(cat "$foreign_lock/owner")" = foreign-owner
test "$(cat "$foreign_lock/volume")" = foreign-volume
if grep -q "rm -f foreign-$operation-lock" "$operation_log"; then
echo "$operation removed a foreign operation lock" >&2
exit 1
fi
start_line=$(grep -n "compose --project-name $project start qdrant" "$operation_log" | sed -n '1s/:.*//p')
release_line=$(grep -n "rm -f $lock_name" "$operation_log" | sed -n '1s/:.*//p')
[ "$start_line" -lt "$release_line" ] || {
echo "$operation released its lock before qdrant reconciliation" >&2
exit 1
}
}
assert_ambiguous_stop_reconciled backup
assert_ambiguous_stop_reconciled restore
echo "qdrant backup/restore archive validation, scoped helper execution, and rollback safety passed."
+1 -1
View File
@@ -85,8 +85,8 @@ validate_volume_metadata "$volume_name"
running_container=$(docker compose --project-name "$project_name" ps --status running -q qdrant)
if [ -n "$running_container" ]; then
docker compose --project-name "$project_name" stop qdrant >/dev/null
restart_qdrant=1
docker compose --project-name "$project_name" stop qdrant >/dev/null
fi
umask 077
+1 -1
View File
@@ -237,8 +237,8 @@ validate_volume_metadata "$volume_name"
running_container=$(docker compose --project-name "$project_name" ps --status running -q qdrant)
if [ -n "$running_container" ]; then
docker compose --project-name "$project_name" stop qdrant >/dev/null
restart_qdrant=1
docker compose --project-name "$project_name" stop qdrant >/dev/null
fi
docker run --rm \