| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | |
| 4 | # Copyright 2026 The HuggingFace Team. All rights reserved. |
| 5 | # ruff: noqa |
| 6 | # fmt: off |
| 7 | # |
| 8 | # Licensed under the Apache License, Version 2.0 (the "License"); |
| 9 | # you may not use this file except in compliance with the License. |
| 10 | # You may obtain a copy of the License at |
| 11 | # |
| 12 | # http://www.apache.org/licenses/LICENSE-2.0 |
| 13 | # |
| 14 | # Unless required by applicable law or agreed to in writing, software |
| 15 | # distributed under the License is distributed on an "AS IS" BASIS, |
| 16 | # WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. |
| 17 | # See the License for the specific language governing permissions and |
| 18 | # limitations under the License. |
| 19 | """This file contains the parsers used by chat response parsing. Each parser takes a chunk of captured text |
| 20 | and parses it into a single key in the output message dictionary. Functions are generally boilerplate.""" |
| 21 | |
| 22 | from __future__ import annotations |
| 23 | |
| 24 | import json |
| 25 | from typing import Any |
| 26 | |
| 27 | import regex as re |
| 28 | |
| 29 | |
| 30 | def _text(text: str, args: dict) -> str: |
| 31 | return text.strip() if args.get("strip", True) else text |
| 32 | |
| 33 | |
| 34 | def _int(text: str, args: dict) -> int: |
| 35 | return int(_text(text, args)) |
| 36 | |
| 37 | |
| 38 | def _float(text: str, args: dict) -> float: |
| 39 | return float(_text(text, args)) |
| 40 | |
| 41 | |
| 42 | def _bool(text: str, args: dict) -> bool: |
| 43 | return _text(text, args).lower() in ("true", "1") |
| 44 | |
| 45 | |
| 46 | # Sentinel characters for lax-JSON string pre-extraction — ASCII control chars |
| 47 | # that should never appear in real LLM output. |
| 48 | _LAX_OPEN, _LAX_CLOSE = "\x01", "\x02" |
| 49 | |
| 50 | |
| 51 | def _json(text: str, args: dict) -> Any: |
| 52 | """JSON parser with optional dialect knobs for LLM-emitted quirks. |
| 53 | |
| 54 | `args`: |
| 55 | - `unquoted_keys` (bool): quote bare-identifier keys before parsing. |
| 56 | - `string_delims` ([[open, close], ...]): strings delimited by these |
| 57 | custom markers are pre-extracted, then restored as standard JSON strings. |
| 58 | - `allow_non_json` (bool): return stripped text if parsing fails. |
| 59 | """ |
| 60 | string_delims = args.get("string_delims", []) |
| 61 | unquoted_keys = args.get("unquoted_keys", False) |
| 62 | |
| 63 | if string_delims and (_LAX_OPEN in text or _LAX_CLOSE in text): |
| 64 | raise ValueError("json: input contains reserved sentinel characters (\\x01/\\x02); cannot parse safely.") |
| 65 | |
| 66 | working = text |
| 67 | captured: list[str] = [] |
| 68 | for open_d, close_d in string_delims: |
| 69 | pattern = re.escape(open_d) + r"(.*?)" + re.escape(close_d) |
| 70 | |
| 71 | def _capture(m: Any) -> str: |
| 72 | captured.append(m.group(1)) |
| 73 | return f"{_LAX_OPEN}{len(captured) - 1}{_LAX_CLOSE}" |
| 74 | |
| 75 | working = re.sub(pattern, _capture, working, flags=re.DOTALL) |
| 76 | |
| 77 | if unquoted_keys: |
| 78 | working = re.sub(r"(?<=[{,])(\w+):", r'"\1":', working) |
| 79 | |
| 80 | for i, s in enumerate(captured): |
| 81 | working = working.replace(f"{_LAX_OPEN}{i}{_LAX_CLOSE}", json.dumps(s)) |
| 82 | |
| 83 | try: |
| 84 | return json.loads(working) |
| 85 | except json.JSONDecodeError as e: |
| 86 | if args.get("allow_non_json"): |
| 87 | return _text(text, args) |
| 88 | if working == text: |
| 89 | raise ValueError(f"json parser could not parse region as JSON.\nContent: {text!r}\nError: {e}") from e |
| 90 | raise ValueError( |
| 91 | f"json: could not parse after dialect transforms.\n" |
| 92 | f"Original: {text!r}\nTransformed: {working!r}\nError: {e}" |
| 93 | ) from e |
| 94 | |
| 95 | |
| 96 | def _sub_parse(raw: str, value_parser: dict | None) -> Any: |
| 97 | if value_parser is None: |
| 98 | return raw |
| 99 | return parse_content(raw, value_parser.get("name", "text"), value_parser.get("args", {})) |
| 100 | |
| 101 | |
| 102 | def _xml_inline(text: str, args: dict) -> dict: |
| 103 | """Parse shallow XML-ish tags into a dict. `tag_pattern` regex must have named |
| 104 | groups `key` and `value`. Optional `value_parser` recurses; `merge_duplicates` |
| 105 | collects duplicate keys into a list.""" |
| 106 | tag_pattern = args.get("tag_pattern") |
| 107 | if tag_pattern is None: |
| 108 | raise ValueError("xml-inline: 'tag_pattern' content_arg is required") |
| 109 | value_parser = args.get("value_parser") |
| 110 | merge = args.get("merge_duplicates", False) |
| 111 | |
| 112 | out: dict[str, Any] = {} |
| 113 | for m in re.finditer(tag_pattern, text, flags=re.DOTALL): |
| 114 | groups = m.groupdict() |
| 115 | key = groups.get("key") |
| 116 | if key is None: |
| 117 | raise ValueError(f"xml-inline: tag_pattern must have a named group 'key'. Pattern: {tag_pattern}") |
| 118 | value = _sub_parse(groups.get("value", ""), value_parser) |
| 119 | if key in out and merge: |
| 120 | if not isinstance(out[key], list): |
| 121 | out[key] = [out[key]] |
| 122 | out[key].append(value) |
| 123 | else: |
| 124 | out[key] = value |
| 125 | return out |
| 126 | |
| 127 | |
| 128 | def _kv_lines(text: str, args: dict) -> dict: |
| 129 | """Parse line-delimited `key<sep>value` pairs into a dict.""" |
| 130 | line_sep = args.get("line_sep", "\n") |
| 131 | kv_sep = args.get("kv_sep", ":") |
| 132 | value_parser = args.get("value_parser") |
| 133 | |
| 134 | out: dict[str, Any] = {} |
| 135 | for line in text.split(line_sep): |
| 136 | line = _text(line, args) |
| 137 | if not line or kv_sep not in line: |
| 138 | continue |
| 139 | k, v = line.split(kv_sep, 1) |
| 140 | k, v = _text(k, args), _text(v, args) |
| 141 | out[k] = _sub_parse(v, value_parser) |
| 142 | return out |
| 143 | |
| 144 | |
| 145 | CONTENT_PARSERS = { |
| 146 | "text": _text, |
| 147 | "int": _int, |
| 148 | "float": _float, |
| 149 | "bool": _bool, |
| 150 | "json": _json, |
| 151 | "xml-inline": _xml_inline, |
| 152 | "kv-lines": _kv_lines, |
| 153 | } |
| 154 | |
| 155 | # Parsers whose output is the verbatim body text (modulo whitespace) — chunks |
| 156 | # from these fields stream with `dirty=False` because each chunk is part of |
| 157 | # the final value. Structured parsers (`json`, `xml-inline`, `kv-lines`) only |
| 158 | # produce a meaningful value on close, so their chunks stream raw bytes |
| 159 | # flagged `dirty=True` while the parsed value is delivered in `region_close`. |
| 160 | STREAMABLE_PARSERS = frozenset({"text", "int", "float", "bool"}) |
| 161 | |
| 162 | |
| 163 | def parse_content(text: str, name: str, args: dict) -> Any: |
| 164 | return CONTENT_PARSERS[name](text, args) |
| 165 | |
| 166 | |
| 167 | _PLACEHOLDER = re.compile(r"\{(\w+(?:\.\w+)*)\}") |
| 168 | |
| 169 | |
| 170 | def _apply_transform(transform: Any, scope: dict) -> Any: |
| 171 | """Recursively walk a transform template, which is used to restructure |
| 172 | parsed output into the actual shape we want. A dotted placeholder like |
| 173 | `{content.args}` descends into keys of the looked-up value.""" |
| 174 | if isinstance(transform, dict): |
| 175 | return {k: _apply_transform(v, scope) for k, v in transform.items()} |
| 176 | if isinstance(transform, list): |
| 177 | return [_apply_transform(v, scope) for v in transform] |
| 178 | if not isinstance(transform, str): |
| 179 | return transform |
| 180 | whole = _PLACEHOLDER.fullmatch(transform) |
| 181 | if not whole: |
| 182 | return transform |
| 183 | path = whole.group(1) |
| 184 | root, *keys = path.split(".") |
| 185 | if root not in scope: |
| 186 | raise KeyError(f"transform placeholder '{{{path}}}' is not defined. Available: {sorted(scope)}") |
| 187 | value = scope[root] |
| 188 | for key in keys: |
| 189 | if not isinstance(value, dict): |
| 190 | raise ValueError(f"transform placeholder '{{{path}}}' cannot index into {type(value).__name__} at '{key}'") |
| 191 | if key not in value: |
| 192 | raise ValueError(f"transform placeholder '{{{path}}}' is missing key '{key}'. Available: {sorted(value)}") |
| 193 | value = value[key] |
| 194 | return value |
| 195 | |
| 196 | |
| 197 | def validate_transform_strings(scope: str, transform: Any) -> None: |
| 198 | """Walk a transform template and reject any string that mixes a `{name}` |
| 199 | placeholder with literal text. Only whole-string placeholders (e.g. |
| 200 | `"{content}"`) and plain literals are supported. Called from the template |
| 201 | loader so authors get a clear error at load time, not at parse time.""" |
| 202 | if isinstance(transform, dict): |
| 203 | for v in transform.values(): |
| 204 | validate_transform_strings(scope, v) |
| 205 | return |
| 206 | if isinstance(transform, list): |
| 207 | for v in transform: |
| 208 | validate_transform_strings(scope, v) |
| 209 | return |
| 210 | if not isinstance(transform, str): |
| 211 | return |
| 212 | if _PLACEHOLDER.search(transform) and not _PLACEHOLDER.fullmatch(transform): |
| 213 | raise ValueError( |
| 214 | f"{scope}: transform string {transform!r} mixes a {{placeholder}} with literal text. " |
| 215 | 'Use either a whole-string placeholder (e.g. "{content}") or a plain literal; ' |
| 216 | "string interpolation is not supported." |
| 217 | ) |
| 218 | |
| 219 | |
| 220 | def process_field(body: str, field, captures: dict) -> Any: |
| 221 | """Run `body` through the field's content parser, then optionally apply the |
| 222 | transform template. When `transform_each` is set, the parsed content must |
| 223 | be a list and the template is applied to each element (with the element's |
| 224 | keys unpacked into the template scope, alongside any regex captures). |
| 225 | |
| 226 | `field` is a `spec.Field`; typed via duck-typing to avoid a cyclic import.""" |
| 227 | value = parse_content(body, field.content, field.content_args) |
| 228 | if field.transform is None: |
| 229 | return value |
| 230 | if field.transform_each: |
| 231 | if not isinstance(value, list): |
| 232 | raise ValueError( |
| 233 | f"Field '{field.name}': transform_each requires the parsed content to be a list, " |
| 234 | f"got {type(value).__name__}." |
| 235 | ) |
| 236 | out = [] |
| 237 | for item in value: |
| 238 | if not isinstance(item, dict): |
| 239 | raise ValueError( |
| 240 | f"Field '{field.name}': transform_each requires each list element to be a dict, " |
| 241 | f"got {type(item).__name__}." |
| 242 | ) |
| 243 | out.append(_apply_transform(field.transform, {**captures, **item})) |
| 244 | return out |
| 245 | return _apply_transform(field.transform, {**captures, "content": value}) |
| 246 | |
| 247 | |
| 248 | __all__ = [ |
| 249 | "CONTENT_PARSERS", |
| 250 | "STREAMABLE_PARSERS", |
| 251 | "parse_content", |
| 252 | "process_field", |
| 253 | "validate_transform_strings", |
| 254 | ] |
| 255 | # fmt: on |