mirror of
https://github.com/bytedance/deer-flow.git
synced 2026-10-26 21:26:39 +00:00
338cd26dc2
* fix(memory): drop unusable numbers in stale extensions and consolidation confidence * fix(memory): log defaulted consolidation confidence and pin the int passthrough
73 lines
3.1 KiB
Python
73 lines
3.1 KiB
Python
"""Tests for non-finite and out-of-range numbers in memory update responses.
|
|
|
|
A model can answer with a JSON number outside float range (``1e400`` parses to
|
|
``inf``) or with ``NaN``, and an unbounded integer such as ``10**400`` raises
|
|
``OverflowError`` in ``float()``. These optional fields are dropped or defaulted
|
|
per entry, as ``_normalize_memory_update_fact`` already does, instead of failing
|
|
the whole batch and losing the valid facts in the same response.
|
|
"""
|
|
|
|
import json
|
|
import logging
|
|
|
|
from deerflow.agents.memory.backends.deermem.deermem.core.updater import _parse_memory_update_response
|
|
|
|
_VALID_FACT = {
|
|
"content": "User prefers dark mode",
|
|
"category": "preference",
|
|
"confidence": 0.9,
|
|
"scope": "user",
|
|
"durability": "durable",
|
|
"authority": "descriptive",
|
|
}
|
|
|
|
|
|
def _response(extra):
|
|
return '{"user":{},"history":{},"newFacts":[' + json.dumps(_VALID_FACT) + "]," + extra + "}"
|
|
|
|
|
|
def test_infinite_extend_by_days_is_dropped():
|
|
# 1e400 is outside float range, so json parses it to inf and int(inf) raises.
|
|
parsed = _parse_memory_update_response(_response('"staleFactsToExtend":[{"id":"fact_1","extend_by_days":1e400}]'))
|
|
|
|
assert len(parsed["newFacts"]) == 1
|
|
assert parsed["staleFactsToExtend"] == []
|
|
|
|
|
|
def test_nan_extend_by_days_is_dropped():
|
|
parsed = _parse_memory_update_response(_response('"staleFactsToExtend":[{"id":"fact_1","extend_by_days":NaN}]'))
|
|
|
|
assert len(parsed["newFacts"]) == 1
|
|
assert parsed["staleFactsToExtend"] == []
|
|
|
|
|
|
def test_unbounded_integer_extend_by_days_is_kept():
|
|
# An integer literal beyond float range must survive as an int: the apply path
|
|
# bounds it with min(days_since + extend_by, staleness_max_extension_days), which
|
|
# works on arbitrary-precision ints but would raise OverflowError on int(float()).
|
|
huge = "1" + "0" * 400
|
|
parsed = _parse_memory_update_response(_response('"staleFactsToExtend":[{"id":"fact_1","extend_by_days":' + huge + "}]"))
|
|
|
|
assert len(parsed["newFacts"]) == 1
|
|
assert parsed["staleFactsToExtend"][0]["extend_by_days"] == 10**400
|
|
|
|
|
|
def test_unbounded_integer_consolidation_confidence_defaults(caplog):
|
|
# float(10**400) raises OverflowError; the entry keeps its default confidence
|
|
# and the default is logged so a systematically mis-scaled model stays visible.
|
|
with caplog.at_level(logging.WARNING):
|
|
parsed = _parse_memory_update_response(_response('"factsToConsolidate":[{"sourceIds":["fact_a","fact_b"],"consolidated":{"content":"Merged fact","category":"knowledge","confidence":' + "1" + "0" * 400 + "}}]"))
|
|
|
|
assert len(parsed["newFacts"]) == 1
|
|
assert parsed["factsToConsolidate"][0]["consolidated"]["confidence"] == 0.9
|
|
assert "out-of-range confidence" in caplog.text
|
|
|
|
|
|
def test_finite_values_are_kept():
|
|
parsed = _parse_memory_update_response(
|
|
_response('"staleFactsToExtend":[{"id":"fact_1","extend_by_days":30}],"factsToConsolidate":[{"sourceIds":["fact_a","fact_b"],"consolidated":{"content":"Merged fact","category":"knowledge","confidence":0.7}}]')
|
|
)
|
|
|
|
assert parsed["staleFactsToExtend"][0]["extend_by_days"] == 30
|
|
assert parsed["factsToConsolidate"][0]["consolidated"]["confidence"] == 0.7
|