Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.
Python agent-contextagent-memoryagent-pluginsagentic-ragcontext-databasedsh-pluginknowledge-basememorymemory-managementself-evolvingvector-database
fix(memory): require parseable final extract output (#1902)
duyua9373665f1d49a
diff --git a/openviking/session/memory/extract_loop.py b/openviking/session/memory/extract_loop.pyindex 9d8beeabe..70fb85e33 100644--- a/openviking/session/memory/extract_loop.py+++ b/openviking/session/memory/extract_loop.py@@ -8,11 +8,16 @@ Reference: bot/vikingbot/agent/loop.py AgentLoop structureimport asyncioimport json-from typing import Any, Dict, List, Optional, Set, Tuple+from typing import Any, Dict, List, Optional, Tuple-from openviking.models.vlm.base import VLMBase, ToolCall+from openviking.models.vlm.base import ToolCall, VLMBasefrom openviking.server.identity import RequestContext-from openviking.session.memory.memory_isolation_handler import RoleScope, MemoryIsolationHandler+from openviking.session.memory.dataclass import (+ MemoryFileContent,+ ResolvedOperation,+ ResolvedOperations,+)+from openviking.session.memory.memory_isolation_handler import MemoryIsolationHandlerfrom openviking.session.memory.schema_model_generator import (SchemaModelGenerator,SchemaPromptGenerator,@@ -20,18 +25,12 @@ from openviking.session.memory.schema_model_generator import (from openviking.session.memory.tools import (MEMORY_TOOLS_REGISTRY,add_tool_call_pair_to_messages,- get_tool,)from openviking.session.memory.utils import (parse_json_with_stability,parse_memory_file_with_fields,pretty_print_messages,)-from openviking.session.memory.dataclass import (- MemoryFileContent,- ResolvedOperation,- ResolvedOperations,-)from openviking.session.memory.utils.json_parser import JsonUtilsfrom openviking.session.memory.utils.uri import supplement_operation_urisfrom openviking.storage.viking_fs import VikingFS, get_viking_fs@@ -93,7 +92,6 @@ class ExtractLoop:self._expected_fields: Optional[List[str]] = Noneself._operations_model: Optional[Any] = None-# Transaction handle for file lockingself._transaction_handle = None# Flag to disable tools in next iteration after unknown tool error@@ -123,7 +121,6 @@ class ExtractLoop:self.schema_prompt_generator = SchemaPromptGenerator(schemas)self.schema_model_generator.generate_all_models()-# 预计算工具 schemasallowed_tools = self.context_provider.get_tools()self._tool_schemas = [@@ -145,12 +142,12 @@ class ExtractLoop:# 预计算 operations_modelrole_scope = self._isolation_handler.get_read_scope() if self._isolation_handler else None- self._operations_model = self.schema_model_generator.create_structured_operations_model(role_scope)+ self._operations_model = self.schema_model_generator.create_structured_operations_model(+ role_scope+ )json_schema = self._operations_model.model_json_schema()--# Build initial messages from providerimport json@@ -189,7 +186,7 @@ The final output of the model must strictly follow the JSON Schema format shownmessages.append({"role": "user",- "content": "You have reached the maximum number of tool call iterations. Do not call any more tools - return your final result directly now.",+ "content": self._build_final_operations_instruction(),})@@ -238,10 +235,12 @@ The final output of the model must strictly follow the JSON Schema format showntracer.info(f"Extended max_iterations to {max_iterations} for format retry")self._add_format_error_message(messages)- # If it's the last iteration, use empty operations+ # If it's the last iteration, fail instead of silently treating an+ # unparseable final response as "no memory operations".if iteration >= max_iterations:- final_operations = ResolvedOperations()- break+ raise RuntimeError(+ "Memory extraction final response could not be parsed as JSON operations"+ )self._disable_tools_for_iteration = Truecontinue@@ -256,9 +255,8 @@ The final output of the model must strictly follow the JSON Schema format shownreturn final_operations, tools_used-async def resolve_operations(self, operations) -> ResolvedOperations:- tracer.info(f'operations={JsonUtils.dumps(operations)}')+ tracer.info(f"operations={JsonUtils.dumps(operations)}")upsert_operations: List[ResolvedOperation] = []delete_file_contents: List[MemoryFileContent] = []errors: List[str] = []@@ -280,7 +278,7 @@ The final output of the model must strictly follow the JSON Schema format shownfor item in items:# 转换为 dictitem_dict = dict(item)- item_dict['memory_type'] = memory_type+ item_dict["memory_type"] = memory_type# 填充 user_id 和 agent_idself._isolation_handler.fill_role_ids(item_dict, role_scope=role_scope)@@ -329,11 +327,8 @@ The final output of the model must strictly follow the JSON Schema format shownop.old_memory_file_content = old_contentbreak--return resolved-@tracer("extract_loop.execute_tool_calls")async def _execute_tool_calls(self, messages, tool_calls, tools_used) -> bool:"""@@ -375,11 +370,6 @@ The final output of the model must strictly follow the JSON Schema format shown})- # Track read tool calls for refetch detection- if tool_call.name == "read" and tool_call.arguments.get("uri"):- uri = tool_call.arguments["uri"]- # 内容由 ToolContext.read_file_contents 记录-add_tool_call_pair_to_messages(messages,call_id=tool_call.id,@@ -390,7 +380,6 @@ The final output of the model must strictly follow the JSON Schema format shownreturn has_unknown_tool-async def _call_llm(self, messages: List[Dict[str, Any]]) -> Tuple[Optional[List], Optional[Any]]:@@ -492,8 +481,6 @@ The final output of the model must strictly follow the JSON Schema format shownprint("No tool calls or operations parsed")return (None, None)--async def _execute_in_parallel(self,tasks: List[Any],@@ -501,10 +488,7 @@ The final output of the model must strictly follow the JSON Schema format shown"""Execute tasks in parallel, similar to AgentLoop."""return await asyncio.gather(*tasks)- async def _check_unread_existing_files(- self,- operations: ResolvedOperations- ) -> Dict:+ async def _check_unread_existing_files(self, operations: ResolvedOperations) -> Dict:refetch_uris = {}for operation in operations.upsert_operations:@@ -513,13 +497,7 @@ The final output of the model must strictly follow the JSON Schema format showncontinuetry:content = await self.context_provider.execute_tool(- ToolCall(- id="",- name="read",- arguments={- "uri": uri- }- )+ ToolCall(id="", name="read", arguments={"uri": uri}))# 读取出错表示文件不存在if isinstance(content, Dict):@@ -544,6 +522,26 @@ The final output of the model must strictly follow the JSON Schema format shown})+ def _build_final_operations_skeleton(self) -> Dict[str, List[Any]]:+ """Build an empty operations object matching the expected flat schema fields."""+ fields = ["delete_uris", *(self._expected_fields or [])]+ return {field: [] for field in dict.fromkeys(fields)}++ def _build_final_operations_instruction(self) -> str:+ """Build schema-aware final-iteration instructions for the LLM."""+ skeleton = json.dumps(+ self._build_final_operations_skeleton(),+ ensure_ascii=False,+ indent=2,+ )+ return (+ "You have reached the maximum number of tool call iterations. "+ "Do not call any more tools. Return your final result now as ONLY a valid JSON object "+ "matching the required schema. Do not include explanations or markdown. "+ "If there are no memory changes, return this exact empty-shape JSON with all fields present:\n"+ f"{skeleton}"+ )+async def _add_refetch_results_to_messages(self,messages: List[Dict[str, Any]],diff --git a/tests/session/memory/test_memory_react.py b/tests/session/memory/test_memory_react.pyindex 740f080b7..3fbeac43c 100644--- a/tests/session/memory/test_memory_react.py+++ b/tests/session/memory/test_memory_react.py@@ -200,3 +200,85 @@ class TestAllowedDirectoriesList:# Verify the result contains the expected directories with variables replacedassert "viking://user/default/memories/preferences" in resultassert "viking://agent/default/memories/tools" in result+++class TestExtractLoopFinalJsonRetry:+ def test_final_instruction_includes_schema_aware_empty_json(self):+ extract_loop = object.__new__(ExtractLoop)+ extract_loop._expected_fields = ["delete_uris", "preferences", "tools"]++ instruction = extract_loop._build_final_operations_instruction()++ assert "ONLY a valid JSON object" in instruction+ assert '"delete_uris": []' in instruction+ assert '"preferences": []' in instruction+ assert '"tools": []' in instruction++ def test_final_skeleton_always_includes_delete_uris(self):+ extract_loop = object.__new__(ExtractLoop)+ extract_loop._expected_fields = ["preferences"]++ assert extract_loop._build_final_operations_skeleton() == {+ "delete_uris": [],+ "preferences": [],+ }++ @pytest.mark.asyncio+ async def test_final_unparseable_response_raises_instead_of_empty_success(self):+ class FakeVLM:+ model = "test-model"++ def __init__(self):+ self.seen_messages = []++ async def get_completion_async(self, **kwargs):+ self.seen_messages.append(list(kwargs["messages"]))+ return "this is not json"++ class FakeContextProvider:+ read_file_contents = {}++ def get_memory_schemas(self, ctx):+ return [+ MemoryTypeSchema(+ memory_type="preferences",+ description="Preferences",+ directory="viking://user/{user_space}/memories/preferences",+ filename_template="{topic}.md",+ fields=[],+ )+ ]++ def get_tools(self):+ return []++ def get_extract_context(self):+ return MagicMock()++ def instruction(self):+ return "Extract memory operations."++ async def prefetch(self):+ return []++ vlm = FakeVLM()+ extract_loop = ExtractLoop(+ vlm=vlm,+ viking_fs=MagicMock(),+ context_provider=FakeContextProvider(),+ max_iterations=1,+ )++ with pytest.raises(RuntimeError, match="final response could not be parsed"):+ await extract_loop.run()++ final_prompts = [+ message["content"]+ for messages in vlm.seen_messages+ for message in messages+ if message.get("role") == "user"+ and "maximum number of tool call iterations" in message.get("content", "")+ ]+ assert final_prompts+ assert '"delete_uris": []' in final_prompts[-1]+ assert '"preferences": []' in final_prompts[-1]