433 lines
15 KiB
Python
433 lines
15 KiB
Python
|
import json
|
||
|
|
from guardrails.types.simple import SimpleTypes
|
||
|
|
import jsonref
|
||
|
|
import regex
|
||
|
|
from typing import Any, Callable, Dict, List, Optional, Set, Tuple, Union, cast
|
||
|
|
|
||
|
|
from guardrails.actions.reask import NonParseableReAsk
|
||
|
|
from guardrails.classes.output_type import OutputTypes
|
||
|
|
from guardrails_ai.types import FailResult
|
||
|
|
from guardrails.schema.parser import get_all_paths
|
||
|
|
from guardrails.utils.safe_get import safe_get
|
||
|
|
|
||
|
|
|
||
|
|
### String to Dictionary Parsing ###
|
||
|
|
def has_code_block(
|
||
|
|
string_value: str, code_type: str = ""
|
||
|
|
) -> Tuple[bool, Optional[int], Optional[int]]:
|
||
|
|
"""Checks if a string contains a code block denoted by leading and trailing
|
||
|
|
tripple ticks (```) with an optional code type for the opening tag.
|
||
|
|
|
||
|
|
Args::
|
||
|
|
string_value (str): The string to check.
|
||
|
|
code_type (str, optional): The specific code type to check for. Defaults to empty string.
|
||
|
|
|
||
|
|
Returns
|
||
|
|
bool: Whether or not the string contains the specified type of code block.
|
||
|
|
int: The starting index of the code block.
|
||
|
|
int: The ending index of the code block.
|
||
|
|
""" # noqa
|
||
|
|
block_border = "```"
|
||
|
|
block_start_border = f"{block_border}{code_type}"
|
||
|
|
|
||
|
|
block_start_index = string_value.find(block_start_border)
|
||
|
|
if block_start_index != -1:
|
||
|
|
block_end_index = string_value.find(
|
||
|
|
block_border, (block_start_index + len(block_start_border))
|
||
|
|
)
|
||
|
|
return (
|
||
|
|
(True, block_start_index, block_end_index)
|
||
|
|
if block_end_index != -1
|
||
|
|
else (False, None, None)
|
||
|
|
)
|
||
|
|
return (False, None, None)
|
||
|
|
|
||
|
|
|
||
|
|
def get_code_block(
|
||
|
|
string_value: str, start: int, end: int, code_type: Optional[str] = ""
|
||
|
|
) -> str:
|
||
|
|
"""Takes a string, start and end indexes, and an optional code type to
|
||
|
|
extract a code block from a string.
|
||
|
|
|
||
|
|
Args::
|
||
|
|
string_value (str): The string to extract a the code block from.
|
||
|
|
start (int): The starting index of the code block. This is assumed to be inclusive of the block boundaries.
|
||
|
|
end (int): The ending index of the code block. This is assumed to be inclusive of the block boundaries.
|
||
|
|
code_type (str, optional): The specific code type to check for. Defaults to empty string.
|
||
|
|
|
||
|
|
Returns:
|
||
|
|
str: The contents of the code block.
|
||
|
|
""" # noqa
|
||
|
|
trimmed_input = string_value
|
||
|
|
|
||
|
|
block_border = "```"
|
||
|
|
block_start_border = f"{block_border}{code_type}"
|
||
|
|
|
||
|
|
start_index = start + len(block_start_border)
|
||
|
|
|
||
|
|
contents = trimmed_input[start_index:end]
|
||
|
|
|
||
|
|
trimmed_output = contents.strip()
|
||
|
|
|
||
|
|
return trimmed_output
|
||
|
|
|
||
|
|
|
||
|
|
def extract_json_from_ouput(
|
||
|
|
output: str,
|
||
|
|
) -> Tuple[Optional[Union[Dict, List]], Optional[Exception]]:
|
||
|
|
# try to load the whole output as json first
|
||
|
|
# there can be corner cases with code blocks
|
||
|
|
# and json/codeblocks inside json
|
||
|
|
try:
|
||
|
|
output_as_dict = json.loads(output, strict=False)
|
||
|
|
return output_as_dict, None
|
||
|
|
except json.decoder.JSONDecodeError:
|
||
|
|
pass
|
||
|
|
|
||
|
|
# Find and extract json from code blocks
|
||
|
|
extracted_code_block = output
|
||
|
|
has_json_block, json_start, json_end = has_code_block(output, "json")
|
||
|
|
if has_json_block and json_start is not None and json_end is not None:
|
||
|
|
extracted_code_block = get_code_block(output, json_start, json_end, "json")
|
||
|
|
else:
|
||
|
|
has_block, block_start, block_end = has_code_block(output)
|
||
|
|
if has_block and block_start is not None and block_end is not None:
|
||
|
|
extracted_code_block = get_code_block(output, block_start, block_end)
|
||
|
|
else:
|
||
|
|
json_pattern = regex.compile(r"\{(?:[^{}]+|\{(?:(?R)|[^{}]+)*\})*\}")
|
||
|
|
json_groups = json_pattern.findall(output)
|
||
|
|
json_start, json_end = output.find("{"), output.rfind("}")
|
||
|
|
if len(json_groups) > 0 and len(json_groups[0]) == (
|
||
|
|
json_end - json_start + 1
|
||
|
|
):
|
||
|
|
extracted_code_block = json_groups[0]
|
||
|
|
|
||
|
|
# Treat the output as a JSON string, and load it into a dict.
|
||
|
|
error = None
|
||
|
|
try:
|
||
|
|
output_as_dict = json.loads(extracted_code_block, strict=False)
|
||
|
|
except json.decoder.JSONDecodeError as e:
|
||
|
|
output_as_dict = None
|
||
|
|
error = e
|
||
|
|
return output_as_dict, error
|
||
|
|
|
||
|
|
|
||
|
|
### Streaming Fragment Parsing ###
|
||
|
|
def is_valid_fragment(fragment: str, verified: set) -> bool:
|
||
|
|
"""Check if the fragment is a somewhat valid JSON."""
|
||
|
|
|
||
|
|
# Strip fragment of whitespaces and newlines
|
||
|
|
# to avoid duplicate checks
|
||
|
|
text = fragment.strip(" \n")
|
||
|
|
|
||
|
|
# Check if text is already verified
|
||
|
|
if text in verified:
|
||
|
|
return False
|
||
|
|
|
||
|
|
# Check if text is valid JSON
|
||
|
|
try:
|
||
|
|
json.loads(text)
|
||
|
|
verified.add(text)
|
||
|
|
return True
|
||
|
|
except ValueError as e:
|
||
|
|
error_msg = str(e)
|
||
|
|
# Check if error is due to missing comma
|
||
|
|
if "Expecting ',' delimiter" in error_msg:
|
||
|
|
verified.add(text)
|
||
|
|
return True
|
||
|
|
return False
|
||
|
|
|
||
|
|
|
||
|
|
def parse_fragment(fragment: str) -> Tuple[Union[str, List, Dict], Optional[str]]:
|
||
|
|
"""Parse the fragment into a dict."""
|
||
|
|
|
||
|
|
# Complete the JSON fragment to handle missing brackets
|
||
|
|
# Stack to keep track of opening brackets
|
||
|
|
stack = []
|
||
|
|
|
||
|
|
# Process each character in the string
|
||
|
|
for char in fragment:
|
||
|
|
if char in "{[":
|
||
|
|
# Push opening brackets onto the stack
|
||
|
|
stack.append(char)
|
||
|
|
elif char in "}]":
|
||
|
|
# Pop from stack if matching opening bracket is found
|
||
|
|
if stack and (
|
||
|
|
(char == "}" and stack[-1] == "{") or (char == "]" and stack[-1] == "[")
|
||
|
|
):
|
||
|
|
stack.pop()
|
||
|
|
|
||
|
|
# Add the necessary closing brackets in reverse order
|
||
|
|
while stack:
|
||
|
|
opening_bracket = stack.pop()
|
||
|
|
if opening_bracket == "{":
|
||
|
|
fragment += "}"
|
||
|
|
elif opening_bracket == "[":
|
||
|
|
fragment += "]"
|
||
|
|
|
||
|
|
# Parse the fragment
|
||
|
|
try:
|
||
|
|
parsed_fragment: Union[Dict, List] = json.loads(fragment)
|
||
|
|
return parsed_fragment, None
|
||
|
|
except ValueError as e:
|
||
|
|
return fragment, str(e)
|
||
|
|
|
||
|
|
|
||
|
|
### LLM Output Parsing ###
|
||
|
|
def parse_json_llm_output(
|
||
|
|
output: str, **kwargs
|
||
|
|
) -> Tuple[
|
||
|
|
Union[str, List, Dict, NonParseableReAsk, None],
|
||
|
|
Union[Optional[Exception], str, bool, None],
|
||
|
|
]:
|
||
|
|
if kwargs.get("stream", False):
|
||
|
|
# Do expected behavior for StreamRunner
|
||
|
|
# 1. Check if the fragment is valid JSON
|
||
|
|
verified = kwargs.get("verified", set())
|
||
|
|
fragment_is_valid = is_valid_fragment(output, verified)
|
||
|
|
if not fragment_is_valid:
|
||
|
|
return output, True
|
||
|
|
|
||
|
|
# 2. Parse the fragment
|
||
|
|
parsed_fragment, parsing_error = parse_fragment(output)
|
||
|
|
return parsed_fragment, parsing_error
|
||
|
|
|
||
|
|
# Else do expected behavior for Runner
|
||
|
|
# Try to get json code block from output.
|
||
|
|
# Return error and reask if it is not parseable.
|
||
|
|
parsed_output, error = extract_json_from_ouput(output)
|
||
|
|
|
||
|
|
if error:
|
||
|
|
reask = NonParseableReAsk(
|
||
|
|
incorrectValue=output,
|
||
|
|
failResults=[
|
||
|
|
FailResult(
|
||
|
|
fixValue=None,
|
||
|
|
errorMessage="Output is not parseable as JSON",
|
||
|
|
)
|
||
|
|
],
|
||
|
|
)
|
||
|
|
return reask, error
|
||
|
|
return parsed_output, None
|
||
|
|
|
||
|
|
|
||
|
|
def parse_string_llm_output(output: str) -> Tuple[str, Optional[Exception]]:
|
||
|
|
# Return a ValueError if the output is empty, else None
|
||
|
|
error = ValueError("Empty response received.") if not output else None
|
||
|
|
return output, error
|
||
|
|
|
||
|
|
|
||
|
|
def parse_llm_output(output: str, output_type: OutputTypes, **kwargs):
|
||
|
|
if output_type == OutputTypes.STRING:
|
||
|
|
return parse_string_llm_output(output)
|
||
|
|
return parse_json_llm_output(output, **kwargs)
|
||
|
|
|
||
|
|
|
||
|
|
def prune_extra_keys(
|
||
|
|
payload: Union[str, List[Any], Dict[str, Any]],
|
||
|
|
schema: Dict[str, Any],
|
||
|
|
*,
|
||
|
|
json_path: str = "$",
|
||
|
|
all_json_paths: Optional[Set[str]] = None,
|
||
|
|
) -> Union[str, List[Any], Dict[str, Any]]:
|
||
|
|
if all_json_paths is None or not len(all_json_paths):
|
||
|
|
all_json_paths = get_all_paths(schema)
|
||
|
|
|
||
|
|
if isinstance(payload, dict):
|
||
|
|
# Do full lookbehind
|
||
|
|
wildcards: List[str] = [
|
||
|
|
path.split(".*")[0] for path in all_json_paths if ".*" in path
|
||
|
|
]
|
||
|
|
ancestor_is_wildcard = any(w in json_path for w in wildcards)
|
||
|
|
actual_keys = list(payload.keys())
|
||
|
|
for key in actual_keys:
|
||
|
|
child_path = f"{json_path}.{key}"
|
||
|
|
if child_path not in all_json_paths and not ancestor_is_wildcard:
|
||
|
|
del payload[key]
|
||
|
|
else:
|
||
|
|
prune_extra_keys(
|
||
|
|
payload=payload.get(key), # type: ignore
|
||
|
|
schema=schema,
|
||
|
|
json_path=child_path,
|
||
|
|
all_json_paths=all_json_paths,
|
||
|
|
)
|
||
|
|
elif isinstance(payload, list):
|
||
|
|
for item in payload:
|
||
|
|
prune_extra_keys(
|
||
|
|
payload=item,
|
||
|
|
schema=schema,
|
||
|
|
json_path=json_path,
|
||
|
|
all_json_paths=all_json_paths,
|
||
|
|
)
|
||
|
|
|
||
|
|
return payload
|
||
|
|
|
||
|
|
|
||
|
|
def coerce(value: Any, desired_type: Callable) -> Any:
|
||
|
|
try:
|
||
|
|
coerced_value = desired_type(value)
|
||
|
|
return coerced_value
|
||
|
|
except (ValueError, TypeError):
|
||
|
|
return value
|
||
|
|
|
||
|
|
|
||
|
|
def try_json_parse(value: str) -> Any:
|
||
|
|
try:
|
||
|
|
return json.loads(value)
|
||
|
|
except Exception:
|
||
|
|
return value
|
||
|
|
|
||
|
|
|
||
|
|
def coerce_to_type(
|
||
|
|
payload: Union[str, List[Any], Dict[str, Any], Any], schema_type: SimpleTypes
|
||
|
|
) -> Any:
|
||
|
|
if schema_type == SimpleTypes.ARRAY:
|
||
|
|
if isinstance(payload, str):
|
||
|
|
payload = try_json_parse(payload)
|
||
|
|
if not isinstance(payload, list):
|
||
|
|
return coerce(payload, list)
|
||
|
|
return payload
|
||
|
|
elif schema_type == SimpleTypes.BOOLEAN:
|
||
|
|
if not isinstance(payload, bool):
|
||
|
|
return coerce(payload, bool)
|
||
|
|
return payload
|
||
|
|
elif schema_type == SimpleTypes.INTEGER:
|
||
|
|
if not isinstance(payload, int):
|
||
|
|
val = coerce(payload, int)
|
||
|
|
return val
|
||
|
|
return payload
|
||
|
|
elif schema_type == SimpleTypes.NULL:
|
||
|
|
return None
|
||
|
|
elif schema_type == SimpleTypes.NUMBER:
|
||
|
|
if not isinstance(payload, float):
|
||
|
|
return coerce(payload, float)
|
||
|
|
return payload
|
||
|
|
elif schema_type == SimpleTypes.OBJECT:
|
||
|
|
if isinstance(payload, str):
|
||
|
|
payload = try_json_parse(payload)
|
||
|
|
if not isinstance(payload, dict):
|
||
|
|
return coerce(payload, dict)
|
||
|
|
return payload
|
||
|
|
elif schema_type == SimpleTypes.STRING:
|
||
|
|
if not isinstance(payload, str) and not isinstance(payload, (list, dict)):
|
||
|
|
return coerce(payload, str)
|
||
|
|
return payload
|
||
|
|
|
||
|
|
|
||
|
|
def coerce_property(
|
||
|
|
payload: Union[str, List[Any], Dict[str, Any], Any], schema: Dict[str, Any]
|
||
|
|
) -> Union[str, List[Any], Dict[str, Any]]:
|
||
|
|
schema_type = schema.get("type")
|
||
|
|
if schema_type:
|
||
|
|
payload = coerce_to_type(payload, schema_type)
|
||
|
|
|
||
|
|
### Schema Composition ###
|
||
|
|
one_of = schema.get("oneOf")
|
||
|
|
if one_of:
|
||
|
|
possible_values = []
|
||
|
|
for sub_schema in one_of:
|
||
|
|
possible_values.append(coerce_property(payload, sub_schema))
|
||
|
|
payload = safe_get(list(filter(None, possible_values)), 0, payload)
|
||
|
|
|
||
|
|
any_of = schema.get("anyOf")
|
||
|
|
if any_of:
|
||
|
|
possible_values = []
|
||
|
|
for sub_schema in any_of:
|
||
|
|
possible_values.append(coerce_property(payload, sub_schema))
|
||
|
|
payload = safe_get(list(filter(None, possible_values)), 0, payload)
|
||
|
|
|
||
|
|
all_of: List[Dict[str, Any]] = schema.get("allOf", [])
|
||
|
|
if all_of:
|
||
|
|
if_blocks = [sub for sub in all_of if sub.get("if")]
|
||
|
|
if if_blocks:
|
||
|
|
for if_block in if_blocks:
|
||
|
|
factored_schema = {**schema, **if_block}
|
||
|
|
factored_schema.pop("allOf", {})
|
||
|
|
payload = coerce_property(payload, factored_schema)
|
||
|
|
|
||
|
|
other_blocks = [sub for sub in all_of if not sub.get("if")]
|
||
|
|
for sub_schema in other_blocks:
|
||
|
|
payload = coerce_property(payload, sub_schema)
|
||
|
|
|
||
|
|
else:
|
||
|
|
factored_schema = {**schema}
|
||
|
|
factored_schema.pop("allOf")
|
||
|
|
for sub_schema in all_of:
|
||
|
|
factored_schema = {**schema, **sub_schema}
|
||
|
|
payload = coerce_property(payload, factored_schema)
|
||
|
|
|
||
|
|
### Object Schema ###
|
||
|
|
properties: Dict[str, Any] = schema.get("properties", {})
|
||
|
|
if properties and isinstance(payload, dict):
|
||
|
|
for k, v in properties.items():
|
||
|
|
payload_value = payload.get(k)
|
||
|
|
if payload_value:
|
||
|
|
payload[k] = coerce_property(payload_value, v)
|
||
|
|
|
||
|
|
### Object Additional Properties ###
|
||
|
|
additional_properties_schema: Dict[str, Any] = schema.get(
|
||
|
|
"additionalProperties", {}
|
||
|
|
)
|
||
|
|
if isinstance(additional_properties_schema, bool):
|
||
|
|
additional_properties_schema = {}
|
||
|
|
if additional_properties_schema and isinstance(payload, dict):
|
||
|
|
declared_properties = properties.keys()
|
||
|
|
additional_properties = [
|
||
|
|
key for key in payload.keys() if key not in declared_properties
|
||
|
|
]
|
||
|
|
for prop in additional_properties:
|
||
|
|
payload_value = payload.get(prop)
|
||
|
|
if payload_value:
|
||
|
|
payload[prop] = coerce_property(
|
||
|
|
payload_value, additional_properties_schema
|
||
|
|
)
|
||
|
|
|
||
|
|
### Conditional SubSchema ###
|
||
|
|
if_block: Dict[str, Any] = schema.get("if", {})
|
||
|
|
if if_block and isinstance(payload, dict):
|
||
|
|
if_properties: Dict[str, Any] = if_block.get("properties", {})
|
||
|
|
|
||
|
|
then_block: Dict[str, Any] = schema.get("then", {})
|
||
|
|
then_properties: Dict[str, Any] = then_block.get("properties", {})
|
||
|
|
|
||
|
|
else_block: Dict[str, Any] = schema.get("else", {})
|
||
|
|
else_properties: Dict[str, Any] = else_block.get("properties", {})
|
||
|
|
|
||
|
|
conditional_schema = else_properties
|
||
|
|
|
||
|
|
condition_satisfied = True
|
||
|
|
for k, v in if_properties.items():
|
||
|
|
actual_value = safe_get(payload, k)
|
||
|
|
condition_value = safe_get(v, "const")
|
||
|
|
condition_satisfied = (
|
||
|
|
condition_satisfied and actual_value == condition_value
|
||
|
|
)
|
||
|
|
|
||
|
|
if condition_satisfied:
|
||
|
|
conditional_schema = then_properties
|
||
|
|
|
||
|
|
factored_schema = {**schema, "properties": {**properties, **conditional_schema}}
|
||
|
|
factored_schema.pop("if", {})
|
||
|
|
factored_schema.pop("then", {})
|
||
|
|
factored_schema.pop("else", {})
|
||
|
|
payload = coerce_property(payload, factored_schema)
|
||
|
|
|
||
|
|
### Array Schema ###
|
||
|
|
item_schema: Dict[str, Any] = schema.get("items", {})
|
||
|
|
if isinstance(payload, list) and item_schema:
|
||
|
|
coerced_items = []
|
||
|
|
for item in payload:
|
||
|
|
coerced_items.append(coerce_property(item, item_schema))
|
||
|
|
payload = coerced_items
|
||
|
|
|
||
|
|
return payload
|
||
|
|
|
||
|
|
|
||
|
|
def coerce_types(
|
||
|
|
payload: Union[str, List[Any], Dict[str, Any], Any], schema: Dict[str, Any]
|
||
|
|
) -> Union[str, List[Any], Dict[str, Any]]:
|
||
|
|
dereferenced_schema = cast(
|
||
|
|
Dict[str, Any], jsonref.replace_refs(schema)
|
||
|
|
) # for pyright
|
||
|
|
return coerce_property(payload, dereferenced_schema)
|