Tolerate common malformed VLM inventory JSON
This commit is contained in:
@@ -45,7 +45,19 @@ def _parse_json(inventory_json: str) -> dict[str, Any]:
|
|||||||
try:
|
try:
|
||||||
value = json.loads(text)
|
value = json.loads(text)
|
||||||
except json.JSONDecodeError as exc:
|
except json.JSONDecodeError as exc:
|
||||||
raise ValueError(f"inventory_json is invalid JSON at line {exc.lineno}, column {exc.colno}: {exc.msg}") from exc
|
repaired = _repair_vlm_json(text)
|
||||||
|
if repaired is None:
|
||||||
|
raise ValueError(
|
||||||
|
f"inventory_json is invalid JSON at line {exc.lineno}, column {exc.colno}: {exc.msg}. "
|
||||||
|
"Check that the VLM returned a complete JSON object with quoted strings."
|
||||||
|
) from exc
|
||||||
|
try:
|
||||||
|
value = json.loads(repaired)
|
||||||
|
except json.JSONDecodeError as repaired_exc:
|
||||||
|
raise ValueError(
|
||||||
|
f"inventory_json is invalid JSON at line {exc.lineno}, column {exc.colno}: {exc.msg}. "
|
||||||
|
f"Automatic VLM repair also failed at line {repaired_exc.lineno}, column {repaired_exc.colno}."
|
||||||
|
) from exc
|
||||||
if not isinstance(value, dict):
|
if not isinstance(value, dict):
|
||||||
raise ValueError("inventory_json must be a JSON object with an 'items' list")
|
raise ValueError("inventory_json must be a JSON object with an 'items' list")
|
||||||
items = value.get("items", [])
|
items = value.get("items", [])
|
||||||
@@ -54,6 +66,81 @@ def _parse_json(inventory_json: str) -> dict[str, Any]:
|
|||||||
return value
|
return value
|
||||||
|
|
||||||
|
|
||||||
|
def _repair_vlm_json(text: str) -> str | None:
|
||||||
|
"""Repair only safe, common VLM JSON damage.
|
||||||
|
|
||||||
|
VLMs frequently emit literal newlines in descriptions, leave a quote
|
||||||
|
unescaped inside prose, or stop before writing the final closing braces.
|
||||||
|
This scanner repairs those cases without changing field names or values.
|
||||||
|
It deliberately does not guess missing commas, keys, or inventory items.
|
||||||
|
"""
|
||||||
|
|
||||||
|
start = text.find("{")
|
||||||
|
if start < 0:
|
||||||
|
return None
|
||||||
|
source = text[start:]
|
||||||
|
output: list[str] = []
|
||||||
|
stack: list[str] = []
|
||||||
|
in_string = False
|
||||||
|
escaped = False
|
||||||
|
root_closed = False
|
||||||
|
index = 0
|
||||||
|
while index < len(source):
|
||||||
|
char = source[index]
|
||||||
|
if root_closed:
|
||||||
|
index += 1
|
||||||
|
continue
|
||||||
|
if in_string:
|
||||||
|
if escaped:
|
||||||
|
output.append(char)
|
||||||
|
escaped = False
|
||||||
|
elif char == "\\":
|
||||||
|
output.append(char)
|
||||||
|
escaped = True
|
||||||
|
elif char == '"':
|
||||||
|
lookahead = index + 1
|
||||||
|
while lookahead < len(source) and source[lookahead].isspace():
|
||||||
|
lookahead += 1
|
||||||
|
# A quote followed by normal prose is part of a VLM string;
|
||||||
|
# JSON string terminators are followed by structure.
|
||||||
|
if lookahead == len(source) or source[lookahead] in ",]}:":
|
||||||
|
output.append(char)
|
||||||
|
in_string = False
|
||||||
|
else:
|
||||||
|
output.append('\\"')
|
||||||
|
elif char in "\r\n":
|
||||||
|
output.append("\\n")
|
||||||
|
elif ord(char) < 32:
|
||||||
|
output.append(f"\\u{ord(char):04x}")
|
||||||
|
else:
|
||||||
|
output.append(char)
|
||||||
|
else:
|
||||||
|
if char == '"':
|
||||||
|
in_string = True
|
||||||
|
output.append(char)
|
||||||
|
elif char in "[{":
|
||||||
|
stack.append(char)
|
||||||
|
output.append(char)
|
||||||
|
elif char in "]}":
|
||||||
|
if stack and ((char == "]" and stack[-1] == "[") or (char == "}" and stack[-1] == "{")):
|
||||||
|
stack.pop()
|
||||||
|
output.append(char)
|
||||||
|
if not stack:
|
||||||
|
root_closed = True
|
||||||
|
else:
|
||||||
|
# Ignore trailing prose or an unmatched closing token.
|
||||||
|
pass
|
||||||
|
else:
|
||||||
|
output.append(char)
|
||||||
|
index += 1
|
||||||
|
if in_string:
|
||||||
|
output.append('"')
|
||||||
|
while stack:
|
||||||
|
output.append("]" if stack.pop() == "[" else "}")
|
||||||
|
repaired = re.sub(r",\s*([}\]])", r"\1", "".join(output)).strip()
|
||||||
|
return repaired if repaired.startswith("{") else None
|
||||||
|
|
||||||
|
|
||||||
def normalize_inventory(inventory_json: str) -> list[dict[str, Any]]:
|
def normalize_inventory(inventory_json: str) -> list[dict[str, Any]]:
|
||||||
"""Parse and validate inventory entries without inventing entries."""
|
"""Parse and validate inventory entries without inventing entries."""
|
||||||
|
|
||||||
|
|||||||
@@ -27,6 +27,20 @@ class ClothingLayoutPlannerTests(unittest.TestCase):
|
|||||||
}) + "\n```")
|
}) + "\n```")
|
||||||
self.assertEqual(items[0]["description"], "blue linen")
|
self.assertEqual(items[0]["description"], "blue linen")
|
||||||
|
|
||||||
|
def test_repairs_literal_newlines_and_unescaped_prose_quotes(self):
|
||||||
|
raw = '{"items":[{"id":"shirt","name":"shirt","description":"Blue shirt\nwith a "soft" collar"}]}'
|
||||||
|
items = PLANNER.normalize_inventory(raw)
|
||||||
|
self.assertEqual(items[0]["description"], 'Blue shirt\nwith a "soft" collar')
|
||||||
|
|
||||||
|
def test_repairs_truncated_vlm_json_by_closing_open_containers(self):
|
||||||
|
raw = '{"items":[{"id":"shoe","name":"shoe","description":"Black leather'
|
||||||
|
items = PLANNER.normalize_inventory(raw)
|
||||||
|
self.assertEqual(items[0]["description"], "Black leather")
|
||||||
|
|
||||||
|
def test_ignores_vlm_prose_after_json_object(self):
|
||||||
|
raw = inventory({"id": "hat", "name": "hat"}) + "\nHere is the inventory."
|
||||||
|
self.assertEqual(PLANNER.normalize_inventory(raw)[0]["id"], "hat")
|
||||||
|
|
||||||
def test_quantities_and_pairs_create_deterministic_instances(self):
|
def test_quantities_and_pairs_create_deterministic_instances(self):
|
||||||
plan = PLANNER.plan_layout(inventory(
|
plan = PLANNER.plan_layout(inventory(
|
||||||
{"id": "sock", "name": "sock", "quantity": 3},
|
{"id": "sock", "name": "sock", "quantity": 3},
|
||||||
|
|||||||
Reference in New Issue
Block a user