Spaces:
Sleeping
Sleeping
File size: 7,955 Bytes
b1fc9c3 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 | """
Model output parser.
Responsible for extracting structured information from the raw text
generated by the LLM:
1. Detect tool calls -- look for ``<tool_call>...</tool_call>`` XML tags
(Qwen style) or bare JSON objects with ``name``/``arguments`` keys.
2. Extract the tool name and argument dict.
3. Validate that the tool name is known and arguments match the schema.
4. Detect a final answer (no tool call present).
5. Handle malformed output via regex fallback and heuristic correction.
"""
from __future__ import annotations
import json
import logging
import re
from dataclasses import dataclass, field
from typing import Any
logger = logging.getLogger(__name__)
# ====================================================================== #
# Parsed result types
# ====================================================================== #
@dataclass
class ParsedAction:
"""Result of parsing a single model response."""
is_tool_call: bool = False
tool_name: str = ""
tool_args: dict[str, Any] = field(default_factory=dict)
thought: str = ""
answer: str = ""
raw: str = ""
parse_error: str | None = None
# ====================================================================== #
# Regex patterns
# ====================================================================== #
# <tool_call>{"name": "...", "arguments": {...}}</tool_call>
_TOOL_CALL_XML_RE = re.compile(
r"<tool_call>\s*(\{.*?\})\s*</tool_call>",
re.DOTALL,
)
# Bare JSON with "name" and "arguments" keys (fallback).
_TOOL_CALL_JSON_RE = re.compile(
r'\{\s*"name"\s*:\s*"(\w+)"\s*,\s*"arguments"\s*:\s*(\{.*?\})\s*\}',
re.DOTALL,
)
# Thought: ... (everything before the tool call or answer).
_THOUGHT_RE = re.compile(
r"(?:Thought|Thinking|Reasoning)\s*:\s*(.+?)(?=\n\s*(?:Action|Answer|<tool_call>|\{)|\Z)",
re.DOTALL | re.IGNORECASE,
)
# Answer: ... or Final Answer: ...
_ANSWER_RE = re.compile(
r"(?:Final\s+)?Answer\s*:\s*(.+)",
re.DOTALL | re.IGNORECASE,
)
# Action: tool_name(arg=value, ...) -- relaxed fallback.
_ACTION_FUNC_RE = re.compile(
r"Action\s*:\s*(\w+)\s*\(([^)]*)\)",
re.IGNORECASE,
)
# ====================================================================== #
# Public API
# ====================================================================== #
def parse_model_output(
text: str,
known_tools: list[str] | None = None,
) -> ParsedAction:
"""Parse a model completion into a structured ``ParsedAction``.
Parameters
----------
text:
Raw model output.
known_tools:
Optional list of registered tool names for validation.
Returns
-------
ParsedAction
"""
result = ParsedAction(raw=text)
# Extract thought.
thought_match = _THOUGHT_RE.search(text)
if thought_match:
result.thought = thought_match.group(1).strip()
# --- Strategy 1: XML-tagged tool call ---
xml_match = _TOOL_CALL_XML_RE.search(text)
if xml_match:
return _parse_json_tool_call(xml_match.group(1), result, known_tools)
# --- Strategy 2: Bare JSON tool call ---
json_match = _TOOL_CALL_JSON_RE.search(text)
if json_match:
name = json_match.group(1)
args_str = json_match.group(2)
try:
args = json.loads(args_str)
except json.JSONDecodeError:
args = _attempt_json_repair(args_str)
if args is not None:
result.is_tool_call = True
result.tool_name = name
result.tool_args = args
return _validate_tool(result, known_tools)
# --- Strategy 3: Action: func(args) fallback ---
func_match = _ACTION_FUNC_RE.search(text)
if func_match:
name = func_match.group(1)
args_raw = func_match.group(2).strip()
args = _parse_kv_args(args_raw)
result.is_tool_call = True
result.tool_name = name
result.tool_args = args
return _validate_tool(result, known_tools)
# --- No tool call detected -- look for an answer ---
answer_match = _ANSWER_RE.search(text)
if answer_match:
result.answer = answer_match.group(1).strip()
return result
# If the model produced text but no recognizable pattern, treat the
# entire output as the answer (common for simple questions).
if text.strip():
# But only if there is no apparent intent to call a tool.
if not any(kw in text.lower() for kw in ("tool_call", "action:", "function")):
result.answer = text.strip()
else:
result.parse_error = (
"Could not parse tool call from model output. "
"Please use the format: <tool_call>{\"name\": \"tool_name\", "
"\"arguments\": {...}}</tool_call>"
)
return result
# ====================================================================== #
# Internal helpers
# ====================================================================== #
def _parse_json_tool_call(
json_str: str,
result: ParsedAction,
known_tools: list[str] | None,
) -> ParsedAction:
"""Parse a JSON tool-call object."""
try:
obj = json.loads(json_str)
except json.JSONDecodeError:
obj = _attempt_json_repair(json_str)
if obj is None:
result.parse_error = f"Malformed JSON in tool_call: {json_str[:200]}"
return result
name = obj.get("name", "")
args = obj.get("arguments", {})
if isinstance(args, str):
try:
args = json.loads(args)
except json.JSONDecodeError:
args = {}
result.is_tool_call = True
result.tool_name = name
result.tool_args = args if isinstance(args, dict) else {}
return _validate_tool(result, known_tools)
def _validate_tool(
result: ParsedAction,
known_tools: list[str] | None,
) -> ParsedAction:
"""Warn if the tool name is not in the known set."""
if known_tools and result.tool_name not in known_tools:
result.parse_error = (
f"Unknown tool '{result.tool_name}'. "
f"Available tools: {known_tools}"
)
return result
def _attempt_json_repair(s: str) -> dict | None:
"""Try common fixes for malformed JSON from LLM output.
Handles: trailing commas, single quotes, unquoted keys.
"""
# Remove trailing commas before } or ].
cleaned = re.sub(r",\s*([}\]])", r"\1", s)
# Replace single quotes with double quotes.
cleaned = cleaned.replace("'", '"')
try:
return json.loads(cleaned)
except json.JSONDecodeError:
pass
# Try wrapping in braces.
if not cleaned.strip().startswith("{"):
try:
return json.loads("{" + cleaned + "}")
except json.JSONDecodeError:
pass
return None
def _parse_kv_args(raw: str) -> dict[str, Any]:
"""Parse ``key=value, key=value`` style arguments.
Used as a last-resort fallback for ``Action: func(key=val, ...)`` patterns.
"""
if not raw:
return {}
args: dict[str, Any] = {}
# Split on commas that are not inside quotes.
parts = re.split(r',\s*(?=\w+=)', raw)
for part in parts:
if "=" not in part:
continue
key, _, val = part.partition("=")
key = key.strip().strip('"').strip("'")
val = val.strip().strip('"').strip("'")
# Attempt type coercion.
if val.lower() in ("true", "false"):
args[key] = val.lower() == "true"
else:
try:
args[key] = int(val)
except ValueError:
try:
args[key] = float(val)
except ValueError:
# Try JSON (for lists / dicts).
try:
args[key] = json.loads(val)
except (json.JSONDecodeError, ValueError):
args[key] = val
return args
|