""" This module will parse the JSON file following the BNF definition: ::= ::= | | ; Where: ; is a valid real number expressed in one of a number of given formats ; is a string of valid characters enclosed in quotes ; is one of the literal strings 'true', 'false', or 'null' (unquoted) ::= | ::= '[' [ *(', ' ) ] ']' ; A sequence of JSON values separated by commas ::= '{' [ *(', ' ) ] '}' ; A sequence of 'members' ::= ': ' ; A pair consisting of a name, and a JSON value If something is wrong (a missing parantheses or quotes for example) it will use a few simple heuristics to fix the JSON string: - Add the missing parentheses if the parser believes that the array or object should be closed - Quote strings or add missing single quotes - Adjust whitespaces and remove line breaks All supported use cases are in the unit tests """ import os import json from typing import Any, Dict, List, Optional, Union, TextIO, Tuple, Literal class StringFileWrapper: # This is a trick to simplify the code, transform the filedescriptor handling into a string handling def __init__(self, fd: TextIO) -> None: self.fd = fd self.length: int = 0 def __getitem__(self, index: Union[int, slice]) -> str: if isinstance(index, slice): self.fd.seek(index.start) value = self.fd.read(index.stop - index.start) self.fd.seek(index.start) return value else: self.fd.seek(index) return self.fd.read(1) def __len__(self) -> int: if self.length < 1: current_position = self.fd.tell() self.fd.seek(0, os.SEEK_END) self.length = self.fd.tell() self.fd.seek(current_position) return self.length class LoggerConfig: # This is a type class to simplify the declaration def __init__(self, log_level: Optional[str]): self.log: List[Dict[str, str]] = [] self.window: int = 10 self.log_level: str = log_level if log_level else "none" JSONReturnType = Union[Dict[str, Any], List[Any], str, float, int, bool, None] class JSONParser: def __init__( self, json_str: Union[str, StringFileWrapper], json_fd: Optional[TextIO], logging: Optional[bool], ) -> None: # The string to parse self.json_str = json_str # Alternatively, the file description with a json file in it if json_fd: # This is a trick we do to treat the file wrapper as an array self.json_str = StringFileWrapper(json_fd) # Index is our iterator that will keep track of which character we are looking at right now self.index: int = 0 # This is used in the object member parsing to manage the special cases of missing quotes in key or value self.context: list[str] = [] # Use this to log the activity, but only if logging is active self.logger = LoggerConfig(log_level="info" if logging else None) def parse( self, ) -> Union[JSONReturnType, Tuple[JSONReturnType, List[Dict[str, str]]]]: json = self.parse_json() if self.index < len(self.json_str): self.log( "The parser returned early, checking if there's more json elements", "info", ) json = [json] last_index = self.index while self.index < len(self.json_str): j = self.parse_json() if j != "": json.append(j) if self.index == last_index: self.index += 1 last_index = self.index # If nothing extra was found, don't return an array if len(json) == 1: self.log( "There were no more elements, returning the element without the array", "info", ) json = json[0] if self.logger.log_level == "none": return json else: return json, self.logger.log def parse_json( self, ) -> JSONReturnType: while True: char = self.get_char_at() # This parser will ignore any basic element (string or number) that is not inside an array or object is_in_context = len(self.context) > 0 # False means that we are at the end of the string provided if char is False: return "" # starts with '{' elif char == "{": self.index += 1 return self.parse_object() # starts with '[' elif char == "[": self.index += 1 return self.parse_array() # there can be an edge case in which a key is empty and at the end of an object # like "key": }. We return an empty string here to close the object properly elif char == "}": self.log( "At the end of an object we found a key with missing value, skipping", "info", ) return "" # starts with a quote elif is_in_context and (char in ['"', "'", "“"] or char.isalpha()): return self.parse_string() # starts with [0-9] or minus elif is_in_context and (char.isdigit() or char == "-" or char == "."): return self.parse_number() # If everything else fails, we just ignore and move on else: self.index += 1 def parse_object(self) -> Dict[str, Any]: # ::= '{' [ *(', ' ) ] '}' ; A sequence of 'members' obj = {} # Stop when you either find the closing parentheses or you have iterated over the entire string while (self.get_char_at() or "}") != "}": # This is what we expect to find: # ::= ': ' # Skip filler whitespaces self.skip_whitespaces_at() # Sometimes LLMs do weird things, if we find a ":" so early, we'll change it to "," and move on if (self.get_char_at() or "") == ":": self.log( "While parsing an object we found a : before a key, ignoring", "info", ) self.index += 1 # We are now searching for they string key # Context is used in the string parser to manage the lack of quotes self.set_context("object_key") self.skip_whitespaces_at() # starts with a key = "" while self.get_char_at(): key = str(self.parse_string()) if key != "" or (key == "" and self.get_char_at() == ":"): # If the string is empty but there is a object divider, we are done here break self.skip_whitespaces_at() # We reached the end here if (self.get_char_at() or "}") == "}": continue self.skip_whitespaces_at() # An extreme case of missing ":" after a key if (self.get_char_at() or "") != ":": self.log( "While parsing an object we missed a : after a key", "info", ) self.index += 1 self.reset_context() self.set_context("object_value") # The value can be any valid json value = self.parse_json() # Reset context since our job is done self.reset_context() obj[key] = value if (self.get_char_at() or "") in [",", "'", '"']: self.index += 1 # Remove trailing spaces self.skip_whitespaces_at() self.index += 1 return obj def parse_array(self) -> List[Any]: # ::= '[' [ *(', ' ) ] ']' ; A sequence of JSON values separated by commas arr = [] self.set_context("array") # Stop when you either find the closing parentheses or you have iterated over the entire string while (self.get_char_at() or "]") != "]": self.skip_whitespaces_at() value = self.parse_json() # It is possible that parse_json() returns nothing valid, so we stop if value == "": break if value == "..." and self.get_char_at(-1) == ".": self.log( "While parsing an array, found a stray '...'; ignoring it", "info" ) else: arr.append(value) # skip over whitespace after a value but before closing ] char = self.get_char_at() while char and (char.isspace() or char == ","): self.index += 1 char = self.get_char_at() # Especially at the end of an LLM generated json you might miss the last "]" char = self.get_char_at() if char and char != "]": self.log( "While parsing an array we missed the closing ], adding it back", "info" ) self.index -= 1 self.index += 1 self.reset_context() return arr def parse_string(self) -> Union[str, bool, None]: # is a string of valid characters enclosed in quotes # i.e. { name: "John" } # Somehow all weird cases in an invalid JSON happen to be resolved in this function, so be careful here # Flag to manage corner cases related to missing starting quote missing_quotes = False doubled_quotes = False lstring_delimiter = rstring_delimiter = '"' char = self.get_char_at() # A valid string can only start with a valid quote or, in our case, with a literal while char and char not in ['"', "'", "“"] and not char.isalnum(): self.index += 1 char = self.get_char_at() if not char: # This is an empty string return "" # Ensuring we use the right delimiter if char == "'": lstring_delimiter = rstring_delimiter = "'" elif char == "“": lstring_delimiter = "“" rstring_delimiter = "”" elif char.isalnum(): # This could be a and not a string. Because (T)rue or (F)alse or (N)ull are valid # But remember, object keys are only of type string if char.lower() in ["t", "f", "n"] and self.get_context() != "object_key": value = self.parse_boolean_or_null() if value != "": return value self.log( "While parsing a string, we found a literal instead of a quote", "info", ) self.log( "While parsing a string, we found no starting quote. Will add the quote back", "info", ) missing_quotes = True if not missing_quotes: self.index += 1 # There is sometimes a weird case of doubled quotes, we manage this also later in the while loop if self.get_char_at() == lstring_delimiter: # If it's an empty key, this was easy if self.get_context() == "object_key" and self.get_char_at(1) == ":": self.index += 1 return "" # Find the next delimiter i = 1 next_c = self.get_char_at(i) while next_c and next_c != rstring_delimiter: i += 1 next_c = self.get_char_at(i) # Now check that the next character is also a delimiter to ensure that we have ""....."" # In that case we ignore this rstring delimiter if next_c and (self.get_char_at(i + 1) or "") == rstring_delimiter: self.log( "While parsing a string, we found a valid starting doubled quote, ignoring it", "info", ) doubled_quotes = True self.index += 1 else: # Ok this is not a doubled quote, check if this is an empty string or not i = 1 next_c = self.get_char_at(i) while next_c and next_c.isspace(): i += 1 next_c = self.get_char_at(i) if next_c not in [",", "]", "}"]: self.log( "While parsing a string, we found a doubled quote but it was a mistake, removing one quote", "info", ) self.index += 1 # Initialize our return value string_acc = "" # Here things get a bit hairy because a string missing the final quote can also be a key or a value in an object # In that case we need to use the ":|,|}" characters as terminators of the string # So this will stop if: # * It finds a closing quote # * It iterated over the entire sequence # * If we are fixing missing quotes in an object, when it finds the special terminators char = self.get_char_at() while char and char != rstring_delimiter: if missing_quotes: if self.get_context() == "object_key" and ( char == ":" or char.isspace() ): self.log( "While parsing a string missing the left delimiter in object key context, we found a :, stopping here", "info", ) break elif self.get_context() == "object_value" and char in [",", "}"]: rstring_delimiter_missing = True # check if this is a case in which the closing comma is NOT missing instead i = 1 next_c = self.get_char_at(i) while next_c and next_c != rstring_delimiter: i += 1 next_c = self.get_char_at(i) if next_c: i += 1 next_c = self.get_char_at(i) # found a delimiter, now we need to check that is followed strictly by a comma or brace while next_c and next_c.isspace(): i += 1 next_c = self.get_char_at(i) if next_c and next_c in [",", "}"]: rstring_delimiter_missing = False if rstring_delimiter_missing: self.log( "While parsing a string missing the left delimiter in object value context, we found a , or } and we couldn't determine that a right delimiter was present. Stopping here", "info", ) break string_acc += char self.index += 1 char = self.get_char_at() if char and len(string_acc) > 0 and string_acc[-1] == "\\": # This is a special case, if people use real strings this might happen self.log("Found a stray escape sequence, normalizing it", "info") string_acc = string_acc[:-1] if char in [rstring_delimiter, "t", "n", "r", "b", "\\"]: escape_seqs = {"t": "\t", "n": "\n", "r": "\r", "b": "\b"} string_acc += escape_seqs.get(char, char) or char self.index += 1 char = self.get_char_at() # ChatGPT sometimes forget to quote stuff in html tags or markdown, so we do this whole thing here if char == rstring_delimiter: # Special case here, in case of double quotes one after another if doubled_quotes and self.get_char_at(1) == rstring_delimiter: self.log( "While parsing a string, we found a doubled quote, ignoring it", "info", ) self.index += 1 elif missing_quotes and self.get_context() == "object_value": # In case of missing starting quote I need to check if the delimeter is the end or the beginning of a key i = 1 next_c = self.get_char_at(i) while next_c and next_c not in [ rstring_delimiter, lstring_delimiter, ]: i += 1 next_c = self.get_char_at(i) if next_c: # We found a quote, now let's make sure there's a ":" following i += 1 next_c = self.get_char_at(i) # found a delimiter, now we need to check that is followed strictly by a comma or brace while next_c and next_c.isspace(): i += 1 next_c = self.get_char_at(i) if next_c and next_c == ":": # Reset the cursor self.index -= 1 char = self.get_char_at() self.log( "In a string with missing quotes and object value context, I found a delimeter but it turns out it was the beginning on the next key. Stopping here.", "info", ) break else: # Check if eventually there is a rstring delimiter, otherwise we bail i = 1 next_c = self.get_char_at(i) check_comma_in_object_value = True while next_c and next_c not in [ rstring_delimiter, lstring_delimiter, ]: # This is a bit of a weird workaround, essentially in object_value context we don't always break on commas # This is because the routine after will make sure to correct any bad guess and this solves a corner case if check_comma_in_object_value and next_c.isalpha(): check_comma_in_object_value = False # If we are in an object context, let's check for the right delimiters if ( ("object_key" in self.context and next_c in [":", "}"]) or ("object_value" in self.context and next_c == "}") or ("array" in self.context and next_c in ["]", ","]) or ( check_comma_in_object_value and self.get_context() == "object_value" and next_c == "," ) ): break i += 1 next_c = self.get_char_at(i) # If we stopped for a comma in object_value context, let's check if find a "} at the end of the string if next_c == "," and self.get_context() == "object_value": i += 1 next_c = self.get_char_at(i) while next_c and next_c != rstring_delimiter: i += 1 next_c = self.get_char_at(i) # Ok now I found a delimiter, let's skip whitespaces and see if next we find a } i += 1 next_c = self.get_char_at(i) while next_c and next_c.isspace(): i += 1 next_c = self.get_char_at(i) if next_c == "}": # OK this is valid then self.log( "While parsing a string, we a misplaced quote that would have closed the string but has a different meaning here since this is the last element of the object, ignoring it", "info", ) string_acc += str(char) self.index += 1 char = self.get_char_at() elif next_c == rstring_delimiter: if self.get_context() == "object_value": # But this might not be it! This could be just a missing comma # We found a delimiter and we need to check if this is a key # so find a rstring_delimiter and a colon after i += 1 next_c = self.get_char_at(i) while next_c and next_c != rstring_delimiter: i += 1 next_c = self.get_char_at(i) i += 1 next_c = self.get_char_at(i) while next_c and next_c != ":": if next_c in [ lstring_delimiter, rstring_delimiter, ",", ]: break i += 1 next_c = self.get_char_at(i) # Only if we fail to find a ':' then we know this is misplaced quote if next_c != ":": self.log( "While parsing a string, we a misplaced quote that would have closed the string but has a different meaning here, ignoring it", "info", ) string_acc += str(char) self.index += 1 char = self.get_char_at() if ( char and missing_quotes and self.get_context() == "object_key" and char.isspace() ): self.log( "While parsing a string, handling an extreme corner case in which the LLM added a comment instead of valid string, invalidate the string and return an empty value", "info", ) self.skip_whitespaces_at() if self.get_char_at() not in [":", ","]: return "" # A fallout of the previous special case in the while loop, # we need to update the index only if we had a closing quote if char != rstring_delimiter: self.log( "While parsing a string, we missed the closing quote, ignoring", "info", ) else: self.index += 1 return string_acc.rstrip() def parse_number(self) -> Union[float, int, str, JSONReturnType]: # is a valid real number expressed in one of a number of given formats number_str = "" number_chars = set("0123456789-.eE/,") char = self.get_char_at() is_array = self.get_context() == "array" while char and char in number_chars and (char != "," or not is_array): number_str += char self.index += 1 char = self.get_char_at() if len(number_str) > 1 and number_str[-1] in "-eE/,": # The number ends with a non valid character for a number/currency, rolling back one number_str = number_str[:-1] self.index -= 1 try: if "," in number_str: return str(number_str) if "." in number_str or "e" in number_str or "E" in number_str: return float(number_str) elif number_str == "-": # If there is a stray "-" this will throw an exception, throw away this character return self.parse_json() else: return int(number_str) except ValueError: return number_str def parse_boolean_or_null(self) -> Union[bool, str, None]: # is one of the literal strings 'true', 'false', or 'null' (unquoted) starting_index = self.index char = (self.get_char_at() or "").lower() value: Optional[Tuple[str, Optional[bool]]] if char == "t": value = ("true", True) elif char == "f": value = ("false", False) elif char == "n": value = ("null", None) if value: i = 0 while char and i < len(value[0]) and char == value[0][i]: i += 1 self.index += 1 char = (self.get_char_at() or "").lower() if i == len(value[0]): return value[1] # If nothing works reset the index before returning self.index = starting_index return "" def get_char_at(self, count: int = 0) -> Union[str, Literal[False]]: # Why not use something simpler? Because try/except in python is a faster alternative to an "if" statement that is often True try: return self.json_str[self.index + count] except IndexError: return False def skip_whitespaces_at(self) -> None: """ This function quickly iterates on whitespaces, syntactic sugar to make the code more concise """ try: char = self.json_str[self.index] except IndexError: return while char.isspace(): self.index += 1 try: char = self.json_str[self.index] except IndexError: return def set_context(self, value: str) -> None: # If a value is provided update the context variable and save in stack if value: self.context.append(value) def reset_context(self) -> None: self.context.pop() def get_context(self) -> str: return self.context[-1] def log(self, text: str, level: str) -> None: if level == self.logger.log_level: context = "" start = max(self.index - self.logger.window, 0) end = min(self.index + self.logger.window, len(self.json_str)) context = self.json_str[start:end] self.logger.log.append( { "text": text, "context": context, } ) def repair_json( json_str: str = "", return_objects: bool = False, skip_json_loads: bool = False, logging: bool = False, json_fd: Optional[TextIO] = None, ensure_ascii: bool = True, ) -> Union[JSONReturnType, Tuple[JSONReturnType, List[Dict[str, str]]]]: """ Given a json formatted string, it will try to decode it and, if it fails, it will try to fix it. It will return the fixed string by default. When `return_objects=True` is passed, it will return the decoded data structure instead. When `skip_json_loads=True` is passed, it will not call the built-in json.loads() function When `logging=True` is passed, it will return a tuple with the repaired json and a log of all repair actions """ parser = JSONParser(json_str, json_fd, logging) if skip_json_loads: parsed_json = parser.parse() else: try: if json_fd: parsed_json = json.load(json_fd) else: parsed_json = json.loads(json_str) except json.JSONDecodeError: parsed_json = parser.parse() # It's useful to return the actual object instead of the json string, # it allows this lib to be a replacement of the json library if return_objects or logging: return parsed_json return json.dumps(parsed_json, ensure_ascii=ensure_ascii) def loads( json_str: str, skip_json_loads: bool = False, logging: bool = False, ) -> Union[JSONReturnType, Tuple[JSONReturnType, List[Dict[str, str]]]]: """ This function works like `json.loads()` except that it will fix your JSON in the process. It is a wrapper around the `repair_json()` function with `return_objects=True`. """ return repair_json( json_str=json_str, return_objects=True, skip_json_loads=skip_json_loads, logging=logging, ) def load( fd: TextIO, skip_json_loads: bool = False, logging: bool = False ) -> Union[JSONReturnType, Tuple[JSONReturnType, List[Dict[str, str]]]]: """ This function works like `json.load()` except that it will fix your JSON in the process. It is a wrapper around the `repair_json()` function with `json_fd=fd` and `return_objects=True`. """ return repair_json( json_fd=fd, return_objects=True, skip_json_loads=skip_json_loads, logging=logging, ) def from_file( filename: str, skip_json_loads: bool = False, logging: bool = False, ) -> Union[JSONReturnType, Tuple[JSONReturnType, List[Dict[str, str]]]]: """ This function is a wrapper around `load()` so you can pass the filename as string """ fd = open(filename) jsonobj = load(fd, skip_json_loads, logging) fd.close() return jsonobj