Skip to content
Merged
Show file tree
Hide file tree
Changes from all commits
Commits
File filter

Filter by extension

Filter by extension

Conversations
Failed to load comments.
Loading
Jump to
Jump to file
Failed to load files.
Loading
Diff view
Diff view
9 changes: 9 additions & 0 deletions src/memos/exceptions.py
Original file line number Diff line number Diff line change
Expand Up @@ -41,3 +41,12 @@ class EmbedderError(MemOSError): ...


class ParserError(MemOSError): ...


class LLMResponseParseError(ParserError):
"""Raised when an LLM response cannot be parsed into the expected shape.

Used in place of a silent ``{}`` return so that callers can distinguish
"the model produced no memories" from "the model produced output we
couldn't parse". Issue #2456.
"""
10 changes: 8 additions & 2 deletions src/memos/mem_reader/multi_modal_struct.py
Original file line number Diff line number Diff line change
Expand Up @@ -10,7 +10,7 @@
from memos import log
from memos.configs.mem_reader import MultiModalStructMemReaderConfig
from memos.context.context import ContextThreadPoolExecutor
from memos.exceptions import EmbedderError
from memos.exceptions import EmbedderError, LLMResponseParseError
from memos.mem_reader.read_multi_modal import MultiModalParser, detect_lang
from memos.mem_reader.read_multi_modal.base import _derive_key
from memos.mem_reader.read_pref_memory.process_preference_memory import process_preference_fine
Expand Down Expand Up @@ -539,6 +539,8 @@ def _get_llm_response(
try:
response_text = self.llm.generate(messages)
response_json = parse_json_result(response_text)
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"[LLM] Exception during chat generation: {e}")
response_json = {
Expand Down Expand Up @@ -732,6 +734,8 @@ def _process_one_item(
# ========== Stage 1: Normal extraction (without reference) ==========
try:
resp = self._get_llm_response(mem_str, custom_tags, sources, prompt_type)
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"[MultiModalFine] Error calling LLM: {e}")
return fine_items
Expand Down Expand Up @@ -876,6 +880,8 @@ def _process_one_item(
result = future.result()
if result:
fine_memory_items.extend(result)
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"[MultiModalFine] worker error: {e} {traceback.format_exc()}")

Expand Down Expand Up @@ -1352,7 +1358,7 @@ def _read_memory(
res_memory = future.result()
if res_memory is not None:
memory_list.append(res_memory)
except EmbedderError:
except (EmbedderError, LLMResponseParseError):
raise
except Exception as e:
logger.error(f"Task failed with exception: {e}")
Expand Down
15 changes: 13 additions & 2 deletions src/memos/mem_reader/simple_struct.py
Original file line number Diff line number Diff line change
Expand Up @@ -15,6 +15,7 @@
from memos.configs.mem_reader import SimpleStructMemReaderConfig
from memos.context.context import ContextThreadPoolExecutor
from memos.embedders.factory import EmbedderFactory
from memos.exceptions import LLMResponseParseError
from memos.llms.factory import LLMFactory
from memos.mem_reader.base import BaseMemReader

Expand Down Expand Up @@ -119,6 +120,8 @@ def _build_node(idx, message, info, source_info, llm, parse_json_result, embedde
if not chunk_res:
logger.warning(f"[Parse] Failed to parse result: {raw}")
return None
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"[Parse] Exception during JSON parsing: {e}")
return None
Expand Down Expand Up @@ -263,10 +266,12 @@ def _safe_generate(self, messages: list[dict]) -> str | None:
return None

def _safe_parse(self, text: str | None) -> dict | None:
if not text:
if text is None:
return None
try:
return parse_json_result(text)
except LLMResponseParseError:
raise
except Exception:
logger.warning("[LLM] JSON parse failed")
return None
Expand All @@ -291,7 +296,7 @@ def _get_llm_response(self, mem_str: str, custom_tags: list[str] | None) -> dict
response_text = self._safe_generate(messages)
response_json = self._safe_parse(response_text)

if not response_json:
if response_json is None:
# NOTE: the key MUST be ``"memory list"`` (with a space) — the
# downstream consumers in ``_process_chat_data`` /
# ``_process_transfer_chat_data`` read via
Expand Down Expand Up @@ -702,6 +707,8 @@ def _read_memory(
res_memory = future.result()
if res_memory is not None:
memory_list.append(res_memory)
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"Task failed with exception: {e}")
logger.error(traceback.format_exc())
Expand Down Expand Up @@ -776,6 +783,8 @@ def fine_transfer_simple_mem(
res_memory = future.result()
if res_memory is not None:
memory_list.append(res_memory)
except LLMResponseParseError:
raise
except Exception as e:
logger.error(f"Task failed with exception: {e}")
logger.error(traceback.format_exc())
Expand Down Expand Up @@ -966,6 +975,8 @@ def _process_doc_data(self, scene_data_info, info, **kwargs):
node = future.result()
if node:
doc_nodes.append(node)
except LLMResponseParseError:
raise
except Exception as e:
tqdm.write(f"[ERROR] {e}")
logger.error(f"[DocReader] Future task failed: {e}")
Expand Down
48 changes: 32 additions & 16 deletions src/memos/mem_reader/utils.py
Original file line number Diff line number Diff line change
Expand Up @@ -2,6 +2,7 @@
import re

from memos import log
from memos.exceptions import LLMResponseParseError


logger = log.get_logger(__name__)
Expand Down Expand Up @@ -36,27 +37,41 @@ def derive_key(text: str, max_len: int = 80) -> str:


def parse_json_result(response_text: str) -> dict:
s = (response_text or "").strip()
"""Parse a JSON object out of an LLM response.

m = re.search(r"```(?:json)?\s*([\s\S]*?)```", s, flags=re.I)
Returns the parsed ``dict`` when the response contains a parseable JSON
object (including a literal ``{}``). Raises ``LLMResponseParseError``
on any other input so callers can distinguish "the model produced no
memories" from "the model produced output we couldn't parse". Issue
#2456.
"""
raw = response_text or ""
s = raw.strip()

if not s:
raise LLMResponseParseError("LLM response was empty")

m = re.search(r"```(?:json)?\s*([\s\S]*?)```", s, flags=re.IGNORECASE)
s = (m.group(1) if m else s.replace("```", "")).strip()

i = s.find("{")
if i == -1:
return {}
raise LLMResponseParseError("No JSON object found in LLM response")
s = s[i:].strip()

last_err: json.JSONDecodeError | None = None

try:
return json.loads(s)
except json.JSONDecodeError:
pass
except json.JSONDecodeError as e:
last_err = e

j = max(s.rfind("}"), s.rfind("]"))
if j != -1:
try:
return json.loads(s[: j + 1])
except json.JSONDecodeError:
pass
except json.JSONDecodeError as e:
last_err = e

def _cheap_close(t: str) -> str:
t += "}" * max(0, t.count("{") - t.count("}"))
Expand All @@ -67,14 +82,15 @@ def _cheap_close(t: str) -> str:
try:
return json.loads(t)
except json.JSONDecodeError as e:
last_err = e
if "Invalid \\escape" in str(e):
s = s.replace("\\", "\\\\")
return json.loads(s)
logger.warning(
f"[JSONParse] Failed to decode JSON: {e}\nTail: Raw {response_text} \
json: {s}"
)
return {}
try:
return json.loads(s.replace("\\", "\\\\"))
except json.JSONDecodeError as e2:
last_err = e2

logger.warning("[JSONParse] Failed to decode JSON at position %s", last_err.pos)
raise LLMResponseParseError("Failed to decode JSON from LLM response") from last_err


def parse_rewritten_response(text: str) -> tuple[bool, dict[int, dict]]:
Expand All @@ -83,7 +99,7 @@ def parse_rewritten_response(text: str) -> tuple[bool, dict[int, dict]]:
Returns (success, parsed_dict) with int keys.
"""
try:
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", text, flags=re.I)
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", text, flags=re.IGNORECASE)
s = (m.group(1) if m else text).strip()
data = json.loads(s)
except Exception:
Expand Down Expand Up @@ -127,7 +143,7 @@ def parse_keep_filter_response(text: str) -> tuple[bool, dict[int, dict]]:
Returns (success, parsed_dict) with int keys.
"""
try:
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", text, flags=re.I)
m = re.search(r"```(?:json)?\s*([\s\S]*?)```", text, flags=re.IGNORECASE)
s = (m.group(1) if m else text).strip()
data = json.loads(s)
except Exception:
Expand Down
Original file line number Diff line number Diff line change
Expand Up @@ -9,6 +9,7 @@
from typing import TYPE_CHECKING

from memos.context.context import ContextThreadPoolExecutor
from memos.exceptions import LLMResponseParseError
from memos.log import get_logger
from memos.mem_scheduler.schemas.message_schemas import ScheduleMessageItem
from memos.mem_scheduler.schemas.task_schemas import (
Expand Down Expand Up @@ -178,6 +179,8 @@ def _process_memories_with_reader(
user_context=user_context,
is_upload_skill=is_upload_skill,
)
except LLMResponseParseError:
raise
except Exception as e:
logger.warning("%s: Fail to transfer mem: %s", e, memory_items, exc_info=True)
return
Expand Down
35 changes: 26 additions & 9 deletions src/memos/memories/textual/tree_text_memory/retrieve/searcher.py
Original file line number Diff line number Diff line change
Expand Up @@ -1167,7 +1167,14 @@ def _sort_and_trim(
include_preference_memory=False,
pref_mem_top_k=6,
):
"""Sort results by score and trim to top_k"""
"""Sort results by score and trim to top_k.

Note:
Never mutates the input items: selected textual items are returned
as independent copies (``model_copy``, no revalidation) carrying the
final ``relativity``; other metadata types are rebuilt into
``SearchedTreeNodeTextualMemoryMetadata`` as before.
"""
final_items = []
if search_tool_memory:
tool_schema_results = [
Expand Down Expand Up @@ -1268,15 +1275,25 @@ def _sort_and_trim(
for item, score in sorted_results:
if plugin and round(score, 2) == 0.00:
continue
meta_data = item.metadata.model_dump()
meta_data["relativity"] = score
final_items.append(
TextualMemoryItem(
id=item.id,
memory=item.memory,
metadata=SearchedTreeNodeTextualMemoryMetadata(**meta_data),
metadata = item.metadata
if isinstance(metadata, SearchedTreeNodeTextualMemoryMetadata):
# Decouple from caller-owned items: deep-copy metadata without
# revalidation (cheap), writing the final score on the copy
# with the same float() coercion the rebuild path applies.
metadata = metadata.model_copy(deep=True, update={"relativity": float(score)})
final_items.append(item.model_copy(update={"metadata": metadata}))
else:
# Type upgrade (e.g. graph/bm25 results without relativity):
# rebuild into the searched metadata type as before.
meta_data = metadata.model_dump()
meta_data["relativity"] = score
final_items.append(
TextualMemoryItem(
id=item.id,
memory=item.memory,
metadata=SearchedTreeNodeTextualMemoryMetadata(**meta_data),
)
)
)
return final_items

@timed
Expand Down
56 changes: 56 additions & 0 deletions tests/api/test_server_router.py
Original file line number Diff line number Diff line change
Expand Up @@ -415,3 +415,59 @@ def test_get_all_response_format(self, mock_handlers, client):
data = response.json()
assert data["message"] == "Memories retrieved successfully"
assert isinstance(data["data"], list)


@pytest.mark.parametrize(
("llm_response", "expected_status"),
[("malformed extraction", 500), ("{}", 200), ('{"memory list": []}', 200)],
)
def test_sync_add_distinguishes_parse_failure_from_empty_result(
mock_init_server, monkeypatch, llm_response, expected_status
):
from types import SimpleNamespace
from unittest.mock import MagicMock

from memos.api.handlers.add_handler import AddHandler
from memos.api.handlers.base_handler import HandlerDependencies
from memos.mem_reader.simple_struct import SimpleStructMemReader

reader = SimpleStructMemReader.__new__(SimpleStructMemReader)
reader.config = SimpleNamespace(remove_prompt_example=False)
reader.llm = MagicMock()
reader.llm.generate.return_value = llm_response
reader.chat_window_max_tokens = 1024
reader._count_tokens = len
reader.save_rawfile = False
reader.memory_version_switch = "off"
cube = MagicMock()
cube.text_mem.add.return_value = []
scheduler = MagicMock()
handler = AddHandler(
HandlerDependencies(
naive_mem_cube=cube,
mem_reader=reader,
mem_scheduler=scheduler,
feedback_server=MagicMock(),
)
)
monkeypatch.setattr("memos.api.routers.server_router.add_handler", handler)

response = TestClient(mock_init_server, raise_server_exceptions=False).post(
"/product/add",
json={
"user_id": "synthetic-user",
"writable_cube_ids": ["synthetic-cube"],
"messages": [{"role": "user", "content": "Synthetic input"}],
"async_mode": "sync",
"mode": "fine",
},
)

assert response.status_code == expected_status
assert response.json()["code"] == expected_status
if expected_status == 500:
cube.text_mem.add.assert_not_called()
scheduler.submit_messages.assert_not_called()
else:
assert response.json()["data"] == []
assert cube.text_mem.add.call_args.args[0] == []
Loading
Loading