Skip to content

Self-evolving Context Database for AI Agents. Unify Agent Memory, Knowledge RAG and Skills.

Feat/mem opt (#1349)

chenjw34f5f6a9ff5f
diff --git a/openviking/client/local.py b/openviking/client/local.py
index a994843a6..874e78054 100644
--- a/openviking/client/local.py
+++ b/openviking/client/local.py
@@ -443,7 +443,7 @@ class LocalClient(BaseClient):
If both content and parts are provided, parts takes precedence.
"""
- from datetime import datetime
+ from datetime import datetime, timezone
from openviking.message.part import Part, TextPart, part_from_dict
@@ -458,15 +458,8 @@ class LocalClient(BaseClient):
else:
raise ValueError("Either content or parts must be provided")
- # 解析 created_at
- msg_created_at = None
- if created_at:
- try:
- msg_created_at = datetime.fromisoformat(created_at)
- except ValueError:
- pass
-
- session.add_message(role, message_parts, created_at=msg_created_at)
+ # created_at 直接传递给 session (毫秒时间戳)
+ session.add_message(role, message_parts, created_at=created_at)
return {
"session_id": session_id,
"message_count": len(session.messages),
diff --git a/openviking/message/message.py b/openviking/message/message.py
index 3280dd06d..51ba13030 100644
--- a/openviking/message/message.py
+++ b/openviking/message/message.py
@@ -21,7 +21,7 @@ class Message:
id: str
role: Literal["user", "assistant"]
parts: List[Part]
- created_at: datetime = None
+ created_at: str = None
@property
def content(self) -> str:
@@ -64,13 +64,12 @@ class Message:
def to_dict(self) -> dict:
"""Serialize to JSONL."""
- created_at_val = self.created_at or datetime.now(timezone.utc)
- created_at_str = format_iso8601(created_at_val)
+ created_at_val = self.created_at or datetime.now(timezone.utc).isoformat()
return {
"id": self.id,
"role": self.role,
"parts": [self._part_to_dict(p) for p in self.parts],
- "created_at": created_at_str,
+ "created_at": created_at_val,
}
def _part_to_dict(self, part: Part) -> dict:
@@ -139,7 +138,7 @@ class Message:
id=data["id"],
role=data["role"],
parts=parts,
- created_at=parse_iso_datetime(data["created_at"]),
+ created_at=data["created_at"],
)
@classmethod
@@ -151,7 +150,7 @@ class Message:
id=msg_id or f"msg_{uuid4().hex}",
role="user",
parts=[TextPart(text=content)],
- created_at=datetime.now(timezone.utc),
+ created_at=datetime.now(timezone.utc).isoformat(),
)
@classmethod
@@ -194,7 +193,7 @@ class Message:
id=msg_id or f"msg_{uuid4().hex}",
role="assistant",
parts=parts,
- created_at=datetime.now(timezone.utc),
+ created_at=datetime.now(timezone.utc).isoformat()
)
def get_context_parts(self) -> List[ContextPart]:
diff --git a/openviking/models/vlm/backends/litellm_vlm.py b/openviking/models/vlm/backends/litellm_vlm.py
index 620085709..e6cca8dc9 100644
--- a/openviking/models/vlm/backends/litellm_vlm.py
+++ b/openviking/models/vlm/backends/litellm_vlm.py
@@ -322,6 +322,7 @@ class LiteLLMVLMProvider(VLMBase):
response = completion(**kwargs)
elapsed = time.perf_counter() - t0
self._update_token_usage_from_response(response, duration_seconds=elapsed)
+ tracer.info(f'response={response}')
if tools:
return self._build_vlm_response(response, has_tools=True)
return self._clean_response(self._extract_content_from_response(response))
@@ -333,7 +334,7 @@ class LiteLLMVLMProvider(VLMBase):
operation_name="LiteLLM VLM completion",
)
- @tracer("vlm.call", ignore_result=False, ignore_args=["messages"])
+ @tracer("litellm.vlm.call", ignore_result=True, ignore_args=["messages"])
async def get_completion_async(
self,
prompt: str = "",
@@ -352,6 +353,7 @@ class LiteLLMVLMProvider(VLMBase):
response = await acompletion(**kwargs)
elapsed = time.perf_counter() - t0
self._update_token_usage_from_response(response, duration_seconds=elapsed)
+ tracer.info(f'response={response}')
if tools:
return self._build_vlm_response(response, has_tools=True)
return self._clean_response(self._extract_content_from_response(response))
diff --git a/openviking/models/vlm/backends/openai_vlm.py b/openviking/models/vlm/backends/openai_vlm.py
index 966686a5a..8668ecd23 100644
--- a/openviking/models/vlm/backends/openai_vlm.py
+++ b/openviking/models/vlm/backends/openai_vlm.py
@@ -350,7 +350,7 @@ class OpenAIVLM(VLMBase):
operation_name="OpenAI VLM completion",
)
- @tracer("vlm.call", ignore_result=True, ignore_args=["messages"])
+ @tracer("openai.vlm.call", ignore_result=True, ignore_args=["messages"])
async def get_completion_async(
self,
prompt: str = "",
diff --git a/openviking/models/vlm/backends/volcengine_vlm.py b/openviking/models/vlm/backends/volcengine_vlm.py
index f5f5370b2..e0b78470e 100644
--- a/openviking/models/vlm/backends/volcengine_vlm.py
+++ b/openviking/models/vlm/backends/volcengine_vlm.py
@@ -130,7 +130,7 @@ class VolcEngineVLM(OpenAIVLM):
return result
return self._clean_response(str(result))
- @tracer("vlm.call")
+ @tracer("volcengine.vlm.call", ignore_result=True, ignore_args=False)
async def get_completion_async(
self,
prompt: str = "",
diff --git a/openviking/prompts/templates/memory/entities.yaml b/openviking/prompts/templates/memory/entities.yaml
index 92f75f1e5..cc5499f78 100644
--- a/openviking/prompts/templates/memory/entities.yaml
+++ b/openviking/prompts/templates/memory/entities.yaml
@@ -2,7 +2,7 @@ memory_type: entities
description: |
Wikipedia article - manages page using Zettelkasten method.
Each page represents an article with relative path links to events.
- Cards should be rich and distributed - avoid putting all info in one card.
+ Entity should be rich and distributed - avoid putting all info in one entity.
directory: "viking://user/{{ user_space }}/memories/entities"
filename_template: "{{ category }}/{{ name }}.md"
enabled: true
@@ -25,10 +25,10 @@ fields:
description: |
- Detailed Zettelkasten card content in markdown format
Relative path format: events://{event_name} or ../events/{year}/{month}/{day}/{event_name}.md
- - Example:
- # LGBTQ+ events Caroline participated in:
- - [Pride parade](../events/2023/03/05/Pride parade.md)
- - [Caroline's school LGBTQ talk](events://Caroline's school LGBTQ talk)
- - [Support group](../events/2024/03/10/Support group.md)
+ - Example:
+ # Book club events Caroline participated in:
+ - [Monthly book discussion](../events/2023/03/05/Monthly book discussion.md)
+ - [Author meetup](events://Author meetup)
+ - [Summer reading challenge](../events/2024/03/10/Summer reading challenge.md)
merge_op: patch
diff --git a/openviking/server/routers/sessions.py b/openviking/server/routers/sessions.py
index d5865b265..919459a69 100644
--- a/openviking/server/routers/sessions.py
+++ b/openviking/server/routers/sessions.py
@@ -259,15 +259,8 @@ async def add_message(
else:
parts = [TextPart(text=request.content or "")]
- # 解析 created_at
- created_at = None
- if request.created_at:
- try:
- created_at = datetime.fromisoformat(request.created_at)
- except ValueError:
- logger.warning(f"Invalid created_at format: {request.created_at}")
-
- session.add_message(request.role, parts, created_at=created_at)
+ # created_at 直接传递给 session (ISO string)
+ session.add_message(request.role, parts, created_at=request.created_at)
return Response(
status="ok",
result={
diff --git a/openviking/session/compressor_v2.py b/openviking/session/compressor_v2.py
index f857d30aa..cc81ec9ef 100644
--- a/openviking/session/compressor_v2.py
+++ b/openviking/session/compressor_v2.py
@@ -14,6 +14,7 @@ from openviking.core.context import Context
from openviking.message import Message
from openviking.server.identity import RequestContext
from openviking.session.memory import ExtractLoop, MemoryUpdater
+from openviking.session.memory.utils.json_parser import JsonUtils
from openviking.storage import VikingDBManager
from openviking.storage.viking_fs import get_viking_fs
from openviking.telemetry import get_current_telemetry
@@ -79,6 +80,7 @@ class SessionCompressorV2:
registry=registry, vikingdb=self.vikingdb, transaction_handle=transaction_handle
)
+ @tracer()
async def extract_long_term_memories(
self,
messages: List[Message],
@@ -93,6 +95,7 @@ class SessionCompressorV2:
Note: Returns empty List[Context] because v2 directly writes to storage.
The list length is used for stats in session.py.
"""
+
if not messages:
return []
@@ -101,8 +104,8 @@ class SessionCompressorV2:
return []
tracer.info("Starting v2 memory extraction from conversation")
+ tracer.info(f"messages={JsonUtils.dumps(messages)}")
config = get_openviking_config()
-
# Initialize default memory files (soul.md, identity.md) if not exist
from openviking.session.memory.memory_type_registry import create_default_registry
diff --git a/openviking/session/memory/extract_loop.py b/openviking/session/memory/extract_loop.py
index 9b35f47eb..296c1b9a5 100644
--- a/openviking/session/memory/extract_loop.py
+++ b/openviking/session/memory/extract_loop.py
@@ -86,6 +86,8 @@ class ExtractLoop:
self._read_files: Set[str] = set()
# Transaction handle for file locking
self._transaction_handle = None
+ # Flag to disable tools in next iteration after unknown tool error
+ self._disable_tools_for_iteration = False
async def run(self) -> Tuple[Optional[Any], List[Dict[str, Any]]]:
"""
@@ -151,7 +153,7 @@ class ExtractLoop:
"role": "system",
"content": f"""
## Output Format
-See the complete JSON Schema below:
+The final output of the model must strictly follow the JSON Schema format shown below:
```json
{schema_str}
```
@@ -169,6 +171,19 @@ See the complete JSON Schema below:
)
messages.extend(tool_call_messages)
+ # Track prefetched files in _read_files to avoid unnecessary refetch
+ for msg in tool_call_messages:
+ if msg.get("role") == "user" and "tool_call_name" in msg.get("content", ""):
+ import json
+ try:
+ content = json.loads(msg.get("content", "{}"))
+ if content.get("tool_call_name") == "read":
+ uri = content.get("args", {}).get("uri")
+ if uri:
+ self._read_files.add(uri)
+ except (json.JSONDecodeError, AttributeError):
+ pass
+
while iteration < max_iterations:
iteration += 1
tracer.info(f"ReAct iteration {iteration}/{max_iterations}")
@@ -187,13 +202,21 @@ See the complete JSON Schema below:
# Call LLM with tools - model decides: tool calls OR final operations
pretty_print_messages(messages)
- tool_calls, operations = await self._call_llm(messages, force_final=is_last_iteration)
+
+ tool_calls, operations = await self._call_llm(
+ messages
+ )
if tool_calls:
- await self._execute_tool_calls(messages, tool_calls, tools_used)
+ has_unknown_tool = await self._execute_tool_calls(messages, tool_calls, tools_used)
+ # If model called an unknown tool, disable tools in next iteration
+ if has_unknown_tool:
+ self._disable_tools_for_iteration = True
+ tracer.info("Unknown tool called, will disable tools in next iteration")
# Allow one extra iteration for refetch
if iteration >= max_iterations:
max_iterations += 1
+ self._disable_tools_for_iteration = True
tracer.info(f"Extended max_iterations to {max_iterations} for tool call")
continue
@@ -222,7 +245,8 @@ See the complete JSON Schema below:
if is_last_iteration:
final_operations = self._operations_model()
break
- # Otherwise continue and try again
+ # Otherwise disable_tools and try again
+ self._disable_tools_for_iteration = True
continue
if final_operations is None:
@@ -236,7 +260,14 @@ See the complete JSON Schema below:
return final_operations, tools_used
@tracer("extract_loop.execute_tool_calls")
- async def _execute_tool_calls(self, messages, tool_calls, tools_used):
+ async def _execute_tool_calls(self, messages, tool_calls, tools_used) -> bool:
+ """
+ Execute tool calls in parallel.
+
+ Returns:
+ True if any tool call returned "Unknown tool" error, indicating
+ the model should not receive tools in the next iteration.
+ """
# Execute all tool calls in parallel
async def execute_single_tool_call(idx: int, tool_call):
"""Execute a single tool call."""
@@ -248,8 +279,13 @@ See the complete JSON Schema below:
]
results = await self._execute_in_parallel(action_tasks)
+ has_unknown_tool = False
+
# Process results and add to messages
for _idx, tool_call, result in results:
+ # Check for unknown tool error
+ if isinstance(result, dict) and result.get("error", "").startswith("Unknown tool:"):
+ has_unknown_tool = True
# Skip if arguments is None
if tool_call.arguments is None:
logger.warning(f"Tool call {tool_call.name} has no arguments, skipping")
@@ -265,7 +301,8 @@ See the complete JSON Schema below:
# Track read tool calls for refetch detection
if tool_call.name == "read" and tool_call.arguments.get("uri"):
- self._read_files.add(tool_call.arguments["uri"])
+ uri = tool_call.arguments["uri"]
+ self._read_files.add(uri)
add_tool_call_pair_to_messages(
messages,
@@ -275,6 +312,8 @@ See the complete JSON Schema below:
result=result,
)
+ return has_unknown_tool
+
def _validate_operations(self, operations: Any) -> None:
"""
Validate that all operations have allowed URIs.
@@ -308,8 +347,7 @@ See the complete JSON Schema below:
async def _call_llm(
self,
- messages: List[Dict[str, Any]],
- force_final: bool = False,
+ messages: List[Dict[str, Any]]
) -> Tuple[Optional[List], Optional[Any]]:
"""
Call LLM with tools. Returns either tool calls OR final operations.
@@ -325,13 +363,17 @@ See the complete JSON Schema below:
await self._mark_cache_breakpoint(messages)
# Call LLM with tools - use tools from strategy
- tool_choice = "none" if force_final else None
-
+ tools = None
+ tool_choice = None
+ if not self._disable_tools_for_iteration:
+ tools = self._tool_schemas
+ tool_choice = "auto"
response = await self.vlm.get_completion_async(
messages=messages,
- tools=self._tool_schemas,
+ tools=tools,
tool_choice=tool_choice,
)
+ tracer.info(f"response={response}")
# print(f'response={response}')
# Log cache hit info
if hasattr(response, "usage") and response.usage:
@@ -352,16 +394,24 @@ See the complete JSON Schema below:
f"[KVCache] prompt_tokens={prompt_tokens}, cached_tokens={cached_tokens}"
)
+ # Case 0: Handle string response (when tools are not provided) or None
+ if response is None:
+ content = ""
+ elif isinstance(response, str):
+ # When tools=None, VLM returns string instead of VLMResponse
+ content = response
# Case 1: LLM returned tool calls
- if response.has_tool_calls:
+ elif response.has_tool_calls:
# Format tool calls nicely for debug logging
for tc in response.tool_calls:
tracer.info(f"[assistant tool_call] (id={tc.id}, name={tc.name})")
tracer.info(f" {json.dumps(tc.arguments, indent=2, ensure_ascii=False)}")
return (response.tool_calls, None)
+ else:
+ # Case 2: VLMResponse without tool calls - get content from response
+ content = response.content or ""
- # Case 2: Try to parse operations from content with stability
- content = response.content or ""
+ # Parse operations from content
if content:
try:
# print(f'LLM response content: {content}')
diff --git a/openviking/session/memory/memory_updater.py b/openviking/session/memory/memory_updater.py
index ad95850b9..15b1f4b2b 100644
--- a/openviking/session/memory/memory_updater.py
+++ b/openviking/session/memory/memory_updater.py
@@ -108,7 +108,12 @@ class ExtractContext:
# elements 可以是 Message 或 str ("...")
elements: List[Message | str] = []
for i, (start, end) in enumerate(ranges):
- if start < 0 or end >= len(self.messages):
+ # 兼容 LLM 提取的 range 越界情况
+ if start < 0:
+ start = 0
+ if end >= len(self.messages):
+ end = len(self.messages) - 1
+ if start > end:
continue
range_msgs = self.messages[start : end + 1]
@@ -140,15 +145,20 @@ class MessageRange:
def _first_message_time(self) -> str | None:
"""获取第一条消息的时间(内部方法)"""
+ from datetime import datetime
+
for elem in self.elements:
if isinstance(elem, str):
continue
if hasattr(elem, "created_at") and elem.created_at:
- return elem.created_at.strftime("%Y-%m-%d")
+ dt = datetime.fromisoformat(elem.created_at)
+ return dt.strftime("%Y-%m-%d")
return None
def _first_message_time_with_weekday(self) -> str | None:
"""获取第一条消息的时间,带周几(内部方法)"""
+ from datetime import datetime
+
for elem in self.elements:
if isinstance(elem, str):
continue
@@ -163,8 +173,9 @@ class MessageRange:
"Saturday",
"Sunday",
]
- weekday = weekday_en[elem.created_at.weekday()]
- return f"{elem.created_at.strftime('%Y-%m-%d')} ({weekday})"
+ dt = datetime.fromisoformat(elem.created_at)
+ weekday = weekday_en[dt.weekday()]
+ return f"{dt.strftime('%Y-%m-%d')} ({weekday})"
return None
diff --git a/openviking/session/memory/session_extract_context_provider.py b/openviking/session/memory/session_extract_context_provider.py
index 87352500a..90be44c83 100644
--- a/openviking/session/memory/session_extract_context_provider.py
+++ b/openviking/session/memory/session_extract_context_provider.py
@@ -18,6 +18,7 @@ from openviking.session.memory.tools import (
get_tool,
)
from openviking.storage.viking_fs import VikingFS
+from openviking.telemetry import tracer
from openviking_cli.utils import get_logger
from openviking_cli.utils.config import get_openviking_config
@@ -104,7 +105,7 @@ The system automatically generates URIs based on memory_type and fields. Just pr
last_msg_time = None
if first_msg_time:
- session_time = first_msg_time
+ session_time = datetime.fromisoformat(first_msg_time)
else:
session_time = datetime.now()
@@ -113,7 +114,8 @@ The system automatically generates URIs based on memory_type and fields. Just pr
# 检查是否需要显示范围
if last_msg_time and last_msg_time != first_msg_time:
- time_display = f"{session_time_str} - {last_msg_time.strftime('%Y-%m-%d %H:%M')}"
+ last_time = datetime.fromisoformat(last_msg_time)
+ time_display = f"{session_time_str} - {last_time.strftime('%Y-%m-%d %H:%M')}"
else:
time_display = session_time_str
diff --git a/openviking/session/memory/tools.py b/openviking/session/memory/tools.py
index 5898dc37c..daf3aa873 100644
--- a/openviking/session/memory/tools.py
+++ b/openviking/session/memory/tools.py
@@ -163,7 +163,7 @@ class MemoryReadTool(MemoryTool):
@property
def description(self) -> str:
- return "Read single file, offset is start line number (0-indexed), limit is number of lines to read, -1 means read to end"
+ return "Read single file"
@property
def parameters(self) -> Dict[str, Any]:
diff --git a/openviking/session/memory/utils/json_parser.py b/openviking/session/memory/utils/json_parser.py
index 71de8b222..6a26522a5 100644
--- a/openviking/session/memory/utils/json_parser.py
+++ b/openviking/session/memory/utils/json_parser.py
@@ -11,6 +11,7 @@ Layer 5: Validation Tolerance - TypeAdapter(strict=False) + list item filtering
"""
import json
+from dataclasses import is_dataclass, asdict
from types import UnionType
from typing import (
Any,
@@ -25,7 +26,8 @@ from typing import (
)
import json_repair
-from pydantic import TypeAdapter
+from pydantic import TypeAdapter, BaseModel, parse_obj_as
+
from openviking_cli.utils import get_logger
@@ -42,9 +44,42 @@ __all__ = [
"_get_origin_type",
"_get_arg_type",
"_any_to_str",
+ "JsonUtils",
]
+
+
+class PydanticEncoder(json.JSONEncoder):
+ def default(self, obj):
+ if isinstance(obj, BaseModel) :
+ # 保存类名和属性值
+ return {
+ **obj.model_dump(mode='python')
+ }
+ elif is_dataclass(obj):
+ return asdict(obj)
+ return super().default(obj)
+
+
+
+class JsonUtils:
+
+ @staticmethod
+ def dumps(obj, indent=4, ensure_ascii=False):
+ if obj is None:
+ return None
+ return json.dumps(obj, ensure_ascii=ensure_ascii, indent=indent, cls=PydanticEncoder)
+
+ @staticmethod
+ def loads(json_str, clazz=None):
+ if not json_str:
+ return None
+ if clazz:
+ return TypeAdapter.validate_python(clazz, json_repair.loads(json_str))
+ return json_repair.loads(json_str)
+
+
def extract_json_content(s: str) -> str:
"""
Layer 1: Extract JSON content from LLM response, removing both leading and trailing non-JSON content.
@@ -440,3 +475,5 @@ def parse_json_with_stability(
return model_class.model_validate(tolerant_data), None
except Exception as e2:
return None, f"Model validation failed even after tolerance: {e} (fallback: {e2})"
+
+
diff --git a/openviking/session/session.py b/openviking/session/session.py
index e3a3d90fc..a4d3830ec 100644
--- a/openviking/session/session.py
+++ b/openviking/session/session.py
@@ -9,7 +9,7 @@ import asyncio
import json
import re
from dataclasses import dataclass, field
-from datetime import datetime
+from datetime import datetime, timezone
from typing import TYPE_CHECKING, Any, Dict, List, Optional
from uuid import uuid4
@@ -169,7 +169,7 @@ class Session:
self.user = user or UserIdentifier.the_default_user()
self.ctx = ctx or RequestContext(user=self.user, role=Role.ROOT)
self.session_id = session_id or str(uuid4())
- self.created_at = datetime.now()
+ self.created_at = int(datetime.now(timezone.utc).timestamp() * 1000)
self._auto_commit_threshold = auto_commit_threshold
self._session_uri = f"viking://session/{self.user.user_space_name()}/{self.session_id}"
@@ -301,14 +301,14 @@ class Session:
self,
role: str,
parts: List[Part],
- created_at: datetime = None,
+ created_at: str = None,
) -> Message:
"""Add a message."""
msg = Message(
id=f"msg_{uuid4().hex}",
role=role,
parts=parts,
- created_at=created_at or datetime.now(),
+ created_at=created_at or datetime.now(timezone.utc).isoformat(),
)
self._messages.append(msg)
@@ -474,7 +474,7 @@ class Session:
"trace_id": trace_id,
}
- @tracer("session_commit_phase2")
+
async def _run_memory_extraction(
self,
task_id: str,
diff --git a/openviking/telemetry/tracer.py b/openviking/telemetry/tracer.py
index 401a10d20..a189fa50d 100644
--- a/openviking/telemetry/tracer.py
+++ b/openviking/telemetry/tracer.py
@@ -472,8 +472,10 @@ class tracer:
if hasattr(current_span, "end_time") and current_span.end_time:
return # span 已结束,不添加 event
current_span.add_event(line)
- except Exception:
- pass
+ except Exception as e:
+
+ import traceback
+ traceback.print_stack()
@staticmethod
def info_span(line: str, console: bool = False) -> None: