Skip to content

RAGFlow is an agent-native database: a Retrieval-Augmented Generation (RAG) engine that grew into context infrastructure.

fix(api): derive SQL retrieval tenant from dataset owner, not chat tenant (#18518)

ChendongPang03d0ad2ace42
diff --git a/api/db/services/dialog_service.py b/api/db/services/dialog_service.py
index 9e0fc05d05..01e3663d04 100644
--- a/api/db/services/dialog_service.py
+++ b/api/db/services/dialog_service.py
@@ -676,21 +676,38 @@ async def async_chat(dialog, messages, stream=True, **kwargs):
logging.debug(f"field_map retrieved: {field_map}")
# try to use sql if field mapping is good to go
if field_map:
- logging.debug("Use SQL to retrieval:{}".format(questions[-1]))
- ans = await use_sql(questions[-1], field_map, dialog.tenant_id, chat_mdl, prompt_config.get("quote", True), dialog.kb_ids, doc_ids=scoped_doc_ids)
- # For aggregate queries (COUNT, SUM, etc.), chunks may be empty but answer is still valid
- if ans and (ans.get("reference", {}).get("chunks") or ans.get("answer")):
- if include_reference_metadata and ans.get("reference", {}).get("chunks"):
- if len(dialog.kb_ids) != 1 and any(not c.get("kb_id") for c in ans["reference"]["chunks"]):
- logging.warning(
- "Skipping some _enrich_chunks_with_document_metadata results because dialog.kb_ids has %d entries and use_sql returned chunks without kb_id.",
- len(dialog.kb_ids),
- )
- _enrich_chunks_with_document_metadata(ans["reference"]["chunks"], metadata_fields)
- yield ans
- return
+ # Derive the doc-store tenant/namespace from the referenced dataset's own
+ # owner, not from dialog.tenant_id: a team-shared dataset may be owned by a
+ # different tenant than the one who created this chat.
+ sql_kbs = [kb for kb in kbs if kb.parser_config and kb.parser_config.get("field_map")]
+ sql_tenant_ids = {kb.tenant_id for kb in sql_kbs}
+ if len(sql_tenant_ids) > 1:
+ # use_sql queries a single tenant's doc-store index per call, and
+ # re-running it once per tenant is too slow to do inline (each call
+ # round-trips an LLM to generate SQL). Skip SQL retrieval rather than
+ # silently querying only one tenant's index and dropping the rest.
+ logging.warning(
+ "Skipping SQL retrieval: field-map datasets span multiple tenants (%s); falling back to vector search.",
+ sql_tenant_ids,
+ )
else:
- logging.debug("SQL failed or returned no results, falling back to vector search")
+ sql_tenant_id = sql_kbs[0].tenant_id if sql_kbs else dialog.tenant_id
+ sql_kb_ids = [kb.id for kb in sql_kbs] if sql_kbs else dialog.kb_ids
+ logging.debug("Use SQL to retrieval:{}".format(questions[-1]))
+ ans = await use_sql(questions[-1], field_map, sql_tenant_id, chat_mdl, prompt_config.get("quote", True), sql_kb_ids, doc_ids=scoped_doc_ids)
+ # For aggregate queries (COUNT, SUM, etc.), chunks may be empty but answer is still valid
+ if ans and (ans.get("reference", {}).get("chunks") or ans.get("answer")):
+ if include_reference_metadata and ans.get("reference", {}).get("chunks"):
+ if len(sql_kb_ids) != 1 and any(not c.get("kb_id") for c in ans["reference"]["chunks"]):
+ logging.warning(
+ "Skipping some _enrich_chunks_with_document_metadata results because sql_kb_ids has %d entries and use_sql returned chunks without kb_id.",
+ len(sql_kb_ids),
+ )
+ _enrich_chunks_with_document_metadata(ans["reference"]["chunks"], metadata_fields)
+ yield ans
+ return
+ else:
+ logging.debug("SQL failed or returned no results, falling back to vector search")
param_keys = [p["key"] for p in prompt_config.get("parameters", [])]
if dialog.kb_ids and "knowledge" not in param_keys and "{knowledge}" in prompt_config.get("system", ""):
diff --git a/test/unit_test/api/db/services/test_dialog_service_use_sql_source_columns.py b/test/unit_test/api/db/services/test_dialog_service_use_sql_source_columns.py
index 7c4c602ffa..dc5f3bef4b 100644
--- a/test/unit_test/api/db/services/test_dialog_service_use_sql_source_columns.py
+++ b/test/unit_test/api/db/services/test_dialog_service_use_sql_source_columns.py
@@ -350,3 +350,247 @@ def test_async_chat_uses_all_docs_when_no_doc_ids_selected(monkeypatch):
assert retriever.calls[0]["kwargs"]["doc_ids"] is None
assert "Chunk text from dataset." in chat_model.calls[0]["system_prompt"]
assert result[0]["answer"] == "stub answer"
+
+
+def test_async_chat_sql_retrieval_uses_dataset_owner_tenant_not_dialog_tenant(monkeypatch):
+ """Regression test for a team-shared dataset SQL-retrieval tenant mismatch.
+
+ When a chat (owned by tenant "chat-creator-tenant") references a table
+ dataset owned by a different tenant ("dataset-owner-tenant"), SQL
+ retrieval must build the doc-store index/table name from the dataset
+ owner's tenant, not from dialog.tenant_id -- otherwise it queries an
+ index that doesn't exist (see GitHub issue #18514).
+ """
+ shared_kb = SimpleNamespace(
+ id="kb-shared",
+ tenant_id="dataset-owner-tenant",
+ parser_config={"field_map": {"product": "Product Name"}},
+ )
+ chat_model = _StubChatModel(["unused"])
+ dialog = SimpleNamespace(
+ kb_ids=["kb-shared"],
+ llm_id="chat-model",
+ tenant_llm_id="",
+ tenant_id="chat-creator-tenant",
+ llm_setting={},
+ similarity_threshold=0.1,
+ vector_similarity_weight=0.2,
+ top_n=8,
+ top_k=32,
+ meta_data_filter=None,
+ prompt_config={
+ "quote": True,
+ "keyword": False,
+ "tts": False,
+ "empty_response": "",
+ "system": "Use only this knowledge: {knowledge}",
+ "parameters": [{"key": "knowledge", "optional": False}],
+ "reasoning": False,
+ "toc_enhance": False,
+ "use_kg": False,
+ },
+ )
+
+ captured_calls = []
+
+ async def _fake_use_sql(_question, _field_map, tenant_id, _chat_mdl, _quote, kb_ids, doc_ids=None):
+ captured_calls.append({"tenant_id": tenant_id, "kb_ids": kb_ids})
+ return {"answer": "sql answer", "reference": {"chunks": [], "doc_aggs": []}, "prompt": ""}
+
+ monkeypatch.setattr(dialog_service, "resolve_model_type", lambda _tid, _llm_id: ["chat"])
+ monkeypatch.setattr(
+ dialog_service,
+ "resolve_model_config",
+ lambda *_args, **_kwargs: {"llm_factory": "unit", "max_tokens": 4096, "model_type": "chat"},
+ )
+ monkeypatch.setattr(dialog_service.TenantLangfuseService, "filter_by_tenant", lambda **_kwargs: None)
+ monkeypatch.setattr(
+ dialog_service,
+ "get_models",
+ lambda _dialog, **_kwargs: ([shared_kb], object(), None, chat_model, None),
+ )
+ monkeypatch.setattr(dialog_service.KnowledgebaseService, "get_field_map", lambda _kb_ids: {"product": "Product Name"})
+ monkeypatch.setattr(dialog_service, "use_sql", _fake_use_sql)
+
+ async def _collect():
+ items = []
+ async for item in dialog_service.async_chat(dialog, [{"role": "user", "content": "How many products?"}], stream=False):
+ items.append(item)
+ return items
+
+ result = asyncio.run(_collect())
+
+ assert len(captured_calls) == 1
+ assert captured_calls[0]["tenant_id"] == "dataset-owner-tenant"
+ assert captured_calls[0]["kb_ids"] == ["kb-shared"]
+ assert result[0]["answer"] == "sql answer"
+
+
+def test_async_chat_sql_retrieval_ignores_empty_field_map_datasets(monkeypatch):
+ mapped_kb = SimpleNamespace(
+ id="kb-mapped",
+ tenant_id="mapped-tenant",
+ parser_config={"field_map": {"product": "Product Name"}},
+ )
+ empty_map_kb = SimpleNamespace(
+ id="kb-empty-map",
+ tenant_id="other-tenant",
+ parser_config={"field_map": {}},
+ )
+ chat_model = _StubChatModel(["unused"])
+ dialog = SimpleNamespace(
+ kb_ids=["kb-mapped", "kb-empty-map"],
+ llm_id="chat-model",
+ tenant_llm_id="",
+ tenant_id="chat-creator-tenant",
+ llm_setting={},
+ similarity_threshold=0.1,
+ vector_similarity_weight=0.2,
+ top_n=8,
+ top_k=32,
+ meta_data_filter=None,
+ prompt_config={
+ "quote": True,
+ "keyword": False,
+ "tts": False,
+ "empty_response": "",
+ "system": "Use only this knowledge: {knowledge}",
+ "parameters": [{"key": "knowledge", "optional": False}],
+ "reasoning": False,
+ "toc_enhance": False,
+ "use_kg": False,
+ },
+ )
+
+ captured_calls = []
+
+ async def _fake_use_sql(_question, _field_map, tenant_id, _chat_mdl, _quote, kb_ids, doc_ids=None):
+ captured_calls.append({"tenant_id": tenant_id, "kb_ids": kb_ids})
+ return {"answer": "sql answer", "reference": {"chunks": [], "doc_aggs": []}, "prompt": ""}
+
+ monkeypatch.setattr(dialog_service, "resolve_model_type", lambda _tid, _llm_id: ["chat"])
+ monkeypatch.setattr(
+ dialog_service,
+ "resolve_model_config",
+ lambda *_args, **_kwargs: {"llm_factory": "unit", "max_tokens": 4096, "model_type": "chat"},
+ )
+ monkeypatch.setattr(dialog_service.TenantLangfuseService, "filter_by_tenant", lambda **_kwargs: None)
+ monkeypatch.setattr(
+ dialog_service,
+ "get_models",
+ lambda _dialog, **_kwargs: ([mapped_kb, empty_map_kb], object(), None, chat_model, None),
+ )
+ monkeypatch.setattr(dialog_service.KnowledgebaseService, "get_field_map", lambda _kb_ids: {"product": "Product Name"})
+ monkeypatch.setattr(dialog_service, "use_sql", _fake_use_sql)
+
+ async def _collect():
+ items = []
+ async for item in dialog_service.async_chat(dialog, [{"role": "user", "content": "How many products?"}], stream=False):
+ items.append(item)
+ return items
+
+ result = asyncio.run(_collect())
+
+ assert captured_calls == [{"tenant_id": "mapped-tenant", "kb_ids": ["kb-mapped"]}]
+ assert result[0]["answer"] == "sql answer"
+
+
+def test_async_chat_skips_sql_retrieval_for_mixed_tenant_field_map_datasets(monkeypatch):
+ """Regression test: SQL retrieval must not silently pick one tenant's index
+ when the field-map datasets referenced by a chat span multiple tenants.
+
+ use_sql() queries a single tenant's doc-store index per call, and re-running
+ it once per tenant inline is too slow (each call round-trips an LLM to
+ generate SQL). So when the datasets span more than one tenant, SQL retrieval
+ is skipped entirely and the flow falls back to vector search, instead of
+ querying only one tenant's index and dropping the other tenants' data.
+ """
+ kb_a = SimpleNamespace(id="kb-a", tenant_id="tenant-a", parser_config={"field_map": {"product": "Product Name"}})
+ kb_b = SimpleNamespace(id="kb-b", tenant_id="tenant-b", parser_config={"field_map": {"product": "Product Name"}})
+
+ retriever = _StubAsyncRetriever(
+ {
+ "total": 1,
+ "chunks": [
+ {
+ "chunk_id": "chunk-1",
+ "content_ltks": "chunk text",
+ "content_with_weight": "Chunk text from dataset.",
+ "doc_id": "doc-1",
+ "docnm_kwd": "doc.txt",
+ "kb_id": "kb-a",
+ "important_kwd": [],
+ "positions": [],
+ "vector": [0.1, 0.2],
+ }
+ ],
+ "doc_aggs": [],
+ }
+ )
+ chat_model = _StubChatModel(["stub answer"])
+ dialog = SimpleNamespace(
+ kb_ids=["kb-a", "kb-b"],
+ llm_id="chat-model",
+ tenant_llm_id="",
+ tenant_id="tenant-a",
+ llm_setting={},
+ similarity_threshold=0.1,
+ vector_similarity_weight=0.2,
+ top_n=8,
+ top_k=32,
+ meta_data_filter=None,
+ prompt_config={
+ "quote": False,
+ "keyword": False,
+ "tts": False,
+ "empty_response": "",
+ "system": "Use only this knowledge: {knowledge}",
+ "parameters": [{"key": "knowledge", "optional": False}],
+ "reasoning": False,
+ "toc_enhance": False,
+ "use_kg": False,
+ },
+ )
+
+ def _unexpected_use_sql(*_args, **_kwargs):
+ raise AssertionError("use_sql must not be called for mixed-tenant field-map datasets")
+
+ monkeypatch.setattr(dialog_service.settings, "retriever", retriever, raising=False)
+ monkeypatch.setattr(dialog_service, "resolve_model_type", lambda _tid, _llm_id: ["chat"])
+ monkeypatch.setattr(
+ dialog_service,
+ "resolve_model_config",
+ lambda *_args, **_kwargs: {"llm_factory": "unit", "max_tokens": 4096, "model_type": "chat"},
+ )
+ monkeypatch.setattr(dialog_service.TenantLangfuseService, "filter_by_tenant", lambda **_kwargs: None)
+ monkeypatch.setattr(
+ dialog_service,
+ "get_models",
+ lambda _dialog, **_kwargs: ([kb_a, kb_b], object(), None, chat_model, None),
+ )
+ monkeypatch.setattr(dialog_service.KnowledgebaseService, "get_field_map", lambda _kb_ids: {"product": "Product Name"})
+ monkeypatch.setattr(dialog_service, "use_sql", _unexpected_use_sql)
+ monkeypatch.setattr(dialog_service, "label_question", lambda _question, _kbs: None)
+ monkeypatch.setattr(
+ dialog_service,
+ "kb_prompt",
+ lambda kbinfos, _max_tokens: ["Chunk text from dataset."] if kbinfos["chunks"] else [],
+ )
+ monkeypatch.setattr(dialog_service, "message_fit_in", lambda msg, _max_tokens: (0, msg))
+
+ async def _collect():
+ items = []
+ async for item in dialog_service.async_chat(dialog, [{"role": "user", "content": "How many products?"}], stream=False):
+ items.append(item)
+ return items
+
+ result = asyncio.run(_collect())
+
+ assert len(retriever.calls) == 1
+ call_args = retriever.calls[0]["args"]
+ assert set(call_args[2]) == {"tenant-a", "tenant-b"}
+ assert call_args[3] == ["kb-a", "kb-b"]
+ assert result[0]["answer"] == "stub answer"
diff --git a/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py b/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
index 8a37ec4aea..a9b4c36ba6 100644
--- a/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
+++ b/test/unit_test/api/db/services/test_gaussdb_dialog_sql.py
@@ -666,7 +666,8 @@ def _run_configured_chat(
kb_id = "abcdefabcdefabcdefabcdefabcdefab"
if embd_mdl is None:
embd_mdl = object()
- get_field_map = Mock(return_value=field_map or {"amount": "number"})
+ resolved_field_map = field_map if field_map is not None else {"amount": "number"}
+ get_field_map = Mock(return_value=resolved_field_map)
monkeypatch.setattr(dialog_service.settings, "retriever", retriever, raising=False)
monkeypatch.setattr(
dialog_service,
@@ -684,7 +685,13 @@ def _run_configured_chat(
monkeypatch.setattr(
dialog_service,
"get_models",
- lambda *_args, **_kwargs: ([types.SimpleNamespace(tenant_id=tenant_id)], embd_mdl, None, chat, None),
+ lambda *_args, **_kwargs: (
+ [types.SimpleNamespace(id=kb_id, tenant_id=tenant_id, parser_config={"field_map": resolved_field_map})],
+ embd_mdl,
+ None,
+ chat,
+ None,
+ ),
)
monkeypatch.setattr(dialog_service, "kb_prompt", lambda *_args, **_kwargs: ["fallback content"])
monkeypatch.setattr(dialog_service, "message_fit_in", lambda messages, _limit: (0, messages))