| 1 | # SPDX-License-Identifier: Apache-2.0 |
| 2 | # SPDX-FileCopyrightText: Copyright contributors to the vLLM project |
| 3 | """Streaming parser engine that orchestrates token ID scanning, |
| 4 | incremental lexing, and state-machine-driven semantic event emission.""" |
| 5 | |
| 6 | from __future__ import annotations |
| 7 | |
| 8 | from collections.abc import Sequence |
| 9 | from dataclasses import dataclass |
| 10 | |
| 11 | from vllm.parser.engine.events import EventType, SemanticEvent |
| 12 | from vllm.parser.engine.incremental_lexer import ( |
| 13 | CONTENT_TERMINAL, |
| 14 | IncrementalLexer, |
| 15 | LexerShape, |
| 16 | LexToken, |
| 17 | TerminalDef, |
| 18 | ) |
| 19 | from vllm.parser.engine.parser_engine_config import ( |
| 20 | ParserEngineConfig, |
| 21 | ParserState, |
| 22 | Transition, |
| 23 | ) |
| 24 | from vllm.parser.engine.token_id_scanner import ( |
| 25 | DROP_TERMINAL, |
| 26 | LexerInput, |
| 27 | PreLexedTerminal, |
| 28 | TextChunk, |
| 29 | TokenIDScanner, |
| 30 | ) |
| 31 | |
| 32 | |
| 33 | @dataclass(slots=True) |
| 34 | class _DropInfo: |
| 35 | lexer_shape: LexerShape |
| 36 | extra_token_ids: dict[int, str] |
| 37 | |
| 38 | |
| 39 | def _build_drop_info( |
| 40 | config: ParserEngineConfig, |
| 41 | tokenizer, |
| 42 | ) -> _DropInfo | None: |
| 43 | try: |
| 44 | special_tokens: list[str] = list(tokenizer.all_special_tokens) |
| 45 | special_ids: list[int] = list(tokenizer.all_special_ids) |
| 46 | except (AttributeError, NotImplementedError): |
| 47 | return None |
| 48 | |
| 49 | if not special_tokens: |
| 50 | return None |
| 51 | |
| 52 | configured_texts = ( |
| 53 | set(config.token_id_terminals.values()) |
| 54 | | config.terminal_literals |
| 55 | | config.preserve_tokens |
| 56 | ) |
| 57 | |
| 58 | extra_token_ids: dict[int, str] = {} |
| 59 | drop_texts: set[str] = set() |
| 60 | for text, tid in zip(special_tokens, special_ids): |
| 61 | if text not in configured_texts: |
| 62 | extra_token_ids[tid] = DROP_TERMINAL |
| 63 | drop_texts.add(text) |
| 64 | |
| 65 | if not drop_texts: |
| 66 | return None |
| 67 | |
| 68 | import regex as re |
| 69 | |
| 70 | drop_terminal_defs = [ |
| 71 | TerminalDef( |
| 72 | name=DROP_TERMINAL, |
| 73 | pattern=re.compile(re.escape(text)), |
| 74 | is_literal=True, |
| 75 | literal=text, |
| 76 | ) |
| 77 | for text in drop_texts |
| 78 | ] |
| 79 | |
| 80 | all_terminal_defs = list(config.terminal_defs) + drop_terminal_defs |
| 81 | lexer_shape = LexerShape(all_terminal_defs) |
| 82 | |
| 83 | return _DropInfo( |
| 84 | lexer_shape=lexer_shape, |
| 85 | extra_token_ids=extra_token_ids, |
| 86 | ) |
| 87 | |
| 88 | |
| 89 | class StreamingParserEngine: |
| 90 | """Consumes ``(delta_text, delta_token_ids)`` pairs and produces a |
| 91 | stream of :class:`SemanticEvent` instances. |
| 92 | |
| 93 | This is the main entry point for streaming parsing. |
| 94 | Create one per request (it is stateful). |
| 95 | |
| 96 | The pipeline is:: |
| 97 | |
| 98 | delta_text + delta_token_ids |
| 99 | → TokenIDScanner (special token pre-lexing) |
| 100 | → IncrementalLexer (text → terminal tokens with prefix buffering) |
| 101 | → State Machine (terminal → semantic events) |
| 102 | → list[SemanticEvent] |
| 103 | |
| 104 | Usage:: |
| 105 | |
| 106 | engine = StreamingParserEngine(config, tokenizer) |
| 107 | for each streaming delta: |
| 108 | events = engine.feed(delta_text, delta_token_ids) |
| 109 | # convert events to DeltaMessage |
| 110 | """ |
| 111 | |
| 112 | def __init__( |
| 113 | self, |
| 114 | config: ParserEngineConfig, |
| 115 | tokenizer, |
| 116 | initial_state: ParserState | None = None, |
| 117 | vocab: dict[str, int] | None = None, |
| 118 | ) -> None: |
| 119 | self.config = config |
| 120 | |
| 121 | resolved_token_ids: dict[int, str] = {} |
| 122 | if tokenizer is not None: |
| 123 | if vocab is None: |
| 124 | vocab = tokenizer.get_vocab() |
| 125 | if config.token_id_terminals: |
| 126 | for terminal_name, token_text in config.token_id_terminals.items(): |
| 127 | tid = vocab.get(token_text) |
| 128 | if tid is not None: |
| 129 | resolved_token_ids[tid] = terminal_name |
| 130 | |
| 131 | drop_info: _DropInfo | None = None |
| 132 | if tokenizer is not None: |
| 133 | drop_info = _build_drop_info(config, tokenizer) |
| 134 | |
| 135 | lexer_shape = config.lexer_shape |
| 136 | if drop_info is not None: |
| 137 | resolved_token_ids.update(drop_info.extra_token_ids) |
| 138 | lexer_shape = drop_info.lexer_shape |
| 139 | |
| 140 | self._resolved_token_ids = resolved_token_ids |
| 141 | self._has_drops = drop_info is not None |
| 142 | |
| 143 | self._scanner = TokenIDScanner( |
| 144 | resolved_token_ids, |
| 145 | tokenizer, |
| 146 | ) |
| 147 | |
| 148 | self._token_id_terminal_names: frozenset[str] = frozenset( |
| 149 | resolved_token_ids.values() |
| 150 | ) |
| 151 | |
| 152 | self._lexer = IncrementalLexer(lexer_shape, content_terminal=CONTENT_TERMINAL) |
| 153 | |
| 154 | self._tool_terminals: frozenset[str] = frozenset( |
| 155 | terminal |
| 156 | for (state, terminal), tr in config.transitions.items() |
| 157 | if tr.next_state in self._TOOL_STATES or state in self._TOOL_STATES |
| 158 | ) |
| 159 | # TOOL_CALL_END may close an inner call rather than its lexical wrapper, |
| 160 | # as in MiniMax, so identify exits from state transitions instead. |
| 161 | self._tool_exit_terminals: frozenset[str] = frozenset( |
| 162 | terminal |
| 163 | for (state, terminal), tr in config.transitions.items() |
| 164 | if state in self._TOOL_STATES and tr.next_state not in self._TOOL_STATES |
| 165 | ) |
| 166 | |
| 167 | self._reasoning_markup_terminals: frozenset[str] = ( |
| 168 | self._compute_reasoning_markup_terminals() |
| 169 | ) |
| 170 | |
| 171 | self.skip_tool_parsing = False |
| 172 | self.skip_reasoning_parsing = False |
| 173 | self.reset(initial_state=initial_state) |
| 174 | |
| 175 | @property |
| 176 | def reasoning_token_count(self) -> int: |
| 177 | return self._reasoning_token_count |
| 178 | |
| 179 | def _record_reasoning_tokens(self, events: Sequence[SemanticEvent]) -> None: |
| 180 | self._reasoning_token_count += sum( |
| 181 | event.token_count |
| 182 | for event in events |
| 183 | if event.type == EventType.REASONING_CHUNK |
| 184 | ) |
| 185 | |
| 186 | def _reset_args_state(self) -> None: |
| 187 | self._args_buffer: str = "" |
| 188 | self._args_safe_end: int = 0 |
| 189 | self._args_brace_depth: int = 0 |
| 190 | self._args_in_string: bool = False |
| 191 | self._args_escape_next: bool = False |
| 192 | |
| 193 | def _reset_array_state(self) -> None: |
| 194 | # State for ``tool_call_body_array``: whether the opening ``[`` has been |
| 195 | # seen, and whether a ``{...}`` element is currently being streamed. |
| 196 | self._array_started: bool = False |
| 197 | self._array_in_element: bool = False |
| 198 | |
| 199 | def reset(self, initial_state: ParserState | None = None) -> None: |
| 200 | """Reset mutable state for reuse across requests. |
| 201 | |
| 202 | Preserves cached immutable structures (compiled terminals, |
| 203 | resolved token IDs, lexer shape, token text cache) to avoid |
| 204 | redundant initialization work. |
| 205 | """ |
| 206 | self.state = ( |
| 207 | initial_state if initial_state is not None else self.config.initial_state |
| 208 | ) |
| 209 | self.tool_index = -1 |
| 210 | self._ever_had_token_ids = False |
| 211 | self._reasoning_token_count = 0 |
| 212 | # DO NOT reset skip_tool_parsing here — callers set it before |
| 213 | # calling methods that trigger reset() (e.g. extract_reasoning), |
| 214 | # and clearing it silently breaks non-streaming tool-call-as- |
| 215 | # implicit-reasoning-end (content returns None). |
| 216 | self._scanner.reset() |
| 217 | self._lexer.reset() |
| 218 | self._message_header_buffer = "" |
| 219 | self._message_header_token_count = 0 |
| 220 | self._in_skipped_tool_span = False |
| 221 | self._reset_args_state() |
| 222 | self._reset_array_state() |
| 223 | |
| 224 | def feed( |
| 225 | self, |
| 226 | delta_text: str, |
| 227 | delta_token_ids: Sequence[int], |
| 228 | ) -> list[SemanticEvent]: |
| 229 | if delta_token_ids: |
| 230 | self._ever_had_token_ids = True |
| 231 | |
| 232 | # Fast path: skip scanner and lexer when the delta is plain |
| 233 | # content with no special tokens and no terminal-starting chars. |
| 234 | if ( |
| 235 | delta_text |
| 236 | and not self._lexer.buffer |
| 237 | and not self._scanner._deferred_terminals |
| 238 | and self._lexer._literal_first_chars.isdisjoint(delta_text) |
| 239 | ): |
| 240 | has_special = False |
| 241 | for tid in delta_token_ids: |
| 242 | if tid in self._resolved_token_ids: |
| 243 | has_special = True |
| 244 | break |
| 245 | if not has_special: |
| 246 | events = self._emit_for_state( |
| 247 | delta_text, token_count=len(delta_token_ids) |
| 248 | ) |
| 249 | self._record_reasoning_tokens(events) |
| 250 | return events |
| 251 | |
| 252 | scanner_items = self._scanner.scan(delta_text, delta_token_ids) |
| 253 | |
| 254 | if len(scanner_items) == 1 and isinstance(scanner_items[0], TextChunk): |
| 255 | item = scanner_items[0] |
| 256 | lex_tokens = self._lexer.feed(item.text, item.token_texts, item.token_count) |
| 257 | if len(lex_tokens) == 1 and lex_tokens[0].terminal == CONTENT_TERMINAL: |
| 258 | events = self._emit_for_state( |
| 259 | lex_tokens[0].value, |
| 260 | token_count=lex_tokens[0].token_count, |
| 261 | ) |
| 262 | else: |
| 263 | events = self._process_lex_tokens(lex_tokens) |
| 264 | self._record_reasoning_tokens(events) |
| 265 | return events |
| 266 | |
| 267 | events = self._process_scanner_items(scanner_items) |
| 268 | self._record_reasoning_tokens(events) |
| 269 | return events |
| 270 | |
| 271 | def _process_scanner_items( |
| 272 | self, items: Sequence[LexerInput] |
| 273 | ) -> list[SemanticEvent]: |
| 274 | events: list[SemanticEvent] = [] |
| 275 | for item in items: |
| 276 | if isinstance(item, PreLexedTerminal): |
| 277 | events.extend(self._process_lex_tokens(self._lexer.flush())) |
| 278 | events.extend(self._on_terminal(item.terminal, item.text)) |
| 279 | elif isinstance(item, TextChunk): |
| 280 | if not item.text and item.token_count: |
| 281 | events.extend( |
| 282 | self._emit_for_state("", token_count=item.token_count) |
| 283 | ) |
| 284 | else: |
| 285 | events.extend( |
| 286 | self._process_lex_tokens( |
| 287 | self._lexer.feed( |
| 288 | item.text, item.token_texts, item.token_count |
| 289 | ) |
| 290 | ) |
| 291 | ) |
| 292 | return events |
| 293 | |
| 294 | def finish(self) -> list[SemanticEvent]: |
| 295 | events = self._process_scanner_items(self._scanner.flush_pending()) |
| 296 | |
| 297 | events.extend(self._process_lex_tokens(self._lexer.flush())) |
| 298 | |
| 299 | if self._args_buffer: |
| 300 | events.append( |
| 301 | SemanticEvent( |
| 302 | EventType.ARG_VALUE_CHUNK, |
| 303 | value=self._args_buffer, |
| 304 | tool_index=self.tool_index, |
| 305 | ) |
| 306 | ) |
| 307 | self._args_buffer = "" |
| 308 | self._args_safe_end = 0 |
| 309 | |
| 310 | if self.state in ( |
| 311 | ParserState.TOOL_PREAMBLE, |
| 312 | ParserState.TOOL_ARGS, |
| 313 | ParserState.TOOL_NAME, |
| 314 | ParserState.TOOL_BETWEEN, |
| 315 | ): |
| 316 | # In array mode a closed element has already emitted its |
| 317 | # TOOL_CALL_END; only a truncated (still-open) element needs one. |
| 318 | end_open_call = ( |
| 319 | self._array_in_element if self.config.tool_call_body_array else True |
| 320 | ) |
| 321 | if end_open_call and self.tool_index >= 0: |
| 322 | events.append( |
| 323 | SemanticEvent( |
| 324 | EventType.TOOL_CALL_END, |
| 325 | tool_index=self.tool_index, |
| 326 | ) |
| 327 | ) |
| 328 | self.state = ParserState.CONTENT |
| 329 | elif self.state == ParserState.REASONING: |
| 330 | events.append( |
| 331 | SemanticEvent(EventType.REASONING_END, tool_index=self.tool_index) |
| 332 | ) |
| 333 | self.state = ParserState.CONTENT |
| 334 | elif self.state == ParserState.MESSAGE_HEADER: |
| 335 | if self._message_header_buffer: |
| 336 | events.append( |
| 337 | SemanticEvent( |
| 338 | EventType.TEXT_CHUNK, |
| 339 | value=self._message_header_buffer, |
| 340 | tool_index=self.tool_index, |
| 341 | token_count=self._message_header_token_count, |
| 342 | ) |
| 343 | ) |
| 344 | self._message_header_buffer = "" |
| 345 | self._message_header_token_count = 0 |
| 346 | self.state = ParserState.CONTENT |
| 347 | |
| 348 | self._record_reasoning_tokens(events) |
| 349 | return events |
| 350 | |
| 351 | def parse_complete(self, text: str) -> list[SemanticEvent]: |
| 352 | token_ids: list[int] = [] |
| 353 | events = self.feed(text, token_ids) |
| 354 | events.extend(self.finish()) |
| 355 | return events |
| 356 | |
| 357 | def _process_lex_tokens(self, tokens: list[LexToken]) -> list[SemanticEvent]: |
| 358 | events: list[SemanticEvent] = [] |
| 359 | strict = self._token_id_terminal_names if self._ever_had_token_ids else None |
| 360 | for tok in tokens: |
| 361 | if tok.terminal == CONTENT_TERMINAL or (strict and tok.terminal in strict): |
| 362 | events.extend(self._on_content(tok.value, tok.token_count)) |
| 363 | else: |
| 364 | events.extend( |
| 365 | self._on_terminal(tok.terminal, tok.value, tok.token_count) |
| 366 | ) |
| 367 | return events |
| 368 | |
| 369 | _TOOL_STATES = frozenset( |
| 370 | { |
| 371 | ParserState.TOOL_PREAMBLE, |
| 372 | ParserState.TOOL_NAME, |
| 373 | ParserState.TOOL_ARGS, |
| 374 | ParserState.TOOL_BETWEEN, |
| 375 | } |
| 376 | ) |
| 377 | |
| 378 | _PLAIN_STATES = frozenset({ParserState.CONTENT, ParserState.REASONING}) |
| 379 | |
| 380 | _REASONING_EVENTS = frozenset({EventType.REASONING_START, EventType.REASONING_END}) |
| 381 | |
| 382 | def _compute_reasoning_markup_terminals(self) -> frozenset[str]: |
| 383 | """Terminals the ``skip_reasoning_parsing`` bypass may neutralize. |
| 384 | |
| 385 | Only reasoning-exclusive markers qualify: every transition they |
| 386 | participate in stays within CONTENT/REASONING and emits nothing |
| 387 | but reasoning events. Inkling's ``<|end_message|>`` is labelled |
| 388 | THINK_END yet also closes text, header, and tool blocks; |
| 389 | bypassing a shared marker would eat that structure, so one impure |
| 390 | marker disables the bypass for the whole config. |
| 391 | """ |
| 392 | markers = frozenset( |
| 393 | terminal |
| 394 | for (state, terminal), tr in self.config.transitions.items() |
| 395 | if ParserState.REASONING in (state, tr.next_state) |
| 396 | and tr.next_state not in self._TOOL_STATES |
| 397 | ) |
| 398 | for (state, terminal), tr in self.config.transitions.items(): |
| 399 | if terminal not in markers: |
| 400 | continue |
| 401 | if ( |
| 402 | state not in self._PLAIN_STATES |
| 403 | or tr.next_state not in self._PLAIN_STATES |
| 404 | or not self._REASONING_EVENTS.issuperset(tr.events) |
| 405 | ): |
| 406 | return frozenset() |
| 407 | return markers |
| 408 | |
| 409 | def _on_terminal( |
| 410 | self, terminal: str, value: str, token_count: int = 0 |
| 411 | ) -> list[SemanticEvent]: |
| 412 | key = (self.state, terminal) |
| 413 | transition = self.config.transitions.get(key) |
| 414 | |
| 415 | if transition is None: |
| 416 | if self._has_drops and terminal == DROP_TERMINAL: |
| 417 | return [] |
| 418 | # The projected skip state may not define the wrapper closer. |
| 419 | if self.skip_tool_parsing and terminal in self._tool_exit_terminals: |
| 420 | self._in_skipped_tool_span = False |
| 421 | return self._emit_for_state(value, token_count) |
| 422 | |
| 423 | if self.skip_reasoning_parsing and terminal in self._reasoning_markup_terminals: |
| 424 | return self._emit_for_state(value, token_count) |
| 425 | |
| 426 | if self.skip_tool_parsing and terminal in self._tool_terminals: |
| 427 | # Inkling reuses one terminal for tool, text, and reasoning exits. |
| 428 | # Outside a forwarded tool span, apply its normal transition. |
| 429 | is_opener = transition.next_state in self._TOOL_STATES |
| 430 | is_exit = terminal in self._tool_exit_terminals |
| 431 | used_as_plain_closer = ( |
| 432 | is_exit and not is_opener and not self._in_skipped_tool_span |
| 433 | ) |
| 434 | if not used_as_plain_closer: |
| 435 | if is_opener: |
| 436 | self._in_skipped_tool_span = True |
| 437 | elif is_exit: |
| 438 | self._in_skipped_tool_span = False |
| 439 | leaving_message_header = self.state == ParserState.MESSAGE_HEADER |
| 440 | if leaving_message_header: |
| 441 | self._message_header_buffer = "" |
| 442 | self._message_header_token_count = 0 |
| 443 | # A tool terminal that implicitly ends reasoning must report |
| 444 | # that even from the header state, or the reasoning pass never |
| 445 | # hands the block to the tool pass. |
| 446 | if EventType.REASONING_END in transition.events: |
| 447 | self.state = ParserState.CONTENT |
| 448 | return [ |
| 449 | SemanticEvent( |
| 450 | EventType.REASONING_END, |
| 451 | value=value, |
| 452 | tool_index=self.tool_index, |
| 453 | ), |
| 454 | SemanticEvent( |
| 455 | EventType.TEXT_CHUNK, |
| 456 | value=value, |
| 457 | tool_index=self.tool_index, |
| 458 | ), |
| 459 | ] |
| 460 | elif leaving_message_header: |
| 461 | self.state = ParserState.CONTENT |
| 462 | return [ |
| 463 | SemanticEvent( |
| 464 | EventType.TEXT_CHUNK, |
| 465 | value=value, |
| 466 | tool_index=self.tool_index, |
| 467 | ) |
| 468 | ] |
| 469 | content_type = self.config.content_events.get(self.state) |
| 470 | if content_type is not None: |
| 471 | return [ |
| 472 | SemanticEvent( |
| 473 | content_type, value=value, tool_index=self.tool_index |
| 474 | ) |
| 475 | ] |
| 476 | return [] |
| 477 | |
| 478 | if transition.skip_in_token_id_mode and self._ever_had_token_ids: |
| 479 | return self._emit_for_state(value, token_count) |
| 480 | |
| 481 | return self._apply_transition(transition, value, token_count) |
| 482 | |
| 483 | def _emit_for_state(self, text: str, token_count: int = 0) -> list[SemanticEvent]: |
| 484 | if self.state == ParserState.MESSAGE_HEADER: |
| 485 | self._message_header_buffer += text |
| 486 | self._message_header_token_count += token_count |
| 487 | return [] |
| 488 | if self.state == ParserState.TOOL_ARGS: |
| 489 | if self.config.tool_call_body_array: |
| 490 | return self._feed_array_text(text) |
| 491 | if self.config.tool_args_json: |
| 492 | return self._feed_args_text(text) |
| 493 | return [ |
| 494 | SemanticEvent( |
| 495 | EventType.ARG_VALUE_CHUNK, |
| 496 | value=text, |
| 497 | tool_index=self.tool_index, |
| 498 | token_count=token_count, |
| 499 | ) |
| 500 | ] |
| 501 | content_type = self.config.content_events.get(self.state) |
| 502 | if content_type is not None: |
| 503 | return [ |
| 504 | SemanticEvent( |
| 505 | content_type, |
| 506 | value=text, |
| 507 | tool_index=self.tool_index, |
| 508 | token_count=token_count, |
| 509 | ) |
| 510 | ] |
| 511 | return [] |
| 512 | |
| 513 | def _on_content(self, text: str, token_count: int = 0) -> list[SemanticEvent]: |
| 514 | if not text: |
| 515 | return [] |
| 516 | return self._emit_for_state(text, token_count) |
| 517 | |
| 518 | def _apply_transition( |
| 519 | self, |
| 520 | transition: Transition, |
| 521 | value: str, |
| 522 | token_count: int = 0, |
| 523 | ) -> list[SemanticEvent]: |
| 524 | events: list[SemanticEvent] = [] |
| 525 | previous_state = self.state |
| 526 | message_header = "" |
| 527 | message_header_token_count = 0 |
| 528 | |
| 529 | if ( |
| 530 | self.state == ParserState.TOOL_ARGS |
| 531 | and transition.next_state != ParserState.TOOL_ARGS |
| 532 | and self._args_buffer |
| 533 | ): |
| 534 | events.append( |
| 535 | SemanticEvent( |
| 536 | EventType.ARG_VALUE_CHUNK, |
| 537 | value=self._args_buffer, |
| 538 | tool_index=self.tool_index, |
| 539 | ) |
| 540 | ) |
| 541 | self._args_buffer = "" |
| 542 | |
| 543 | if previous_state == ParserState.MESSAGE_HEADER: |
| 544 | message_header = self._message_header_buffer |
| 545 | message_header_token_count = self._message_header_token_count |
| 546 | self._message_header_buffer = "" |
| 547 | self._message_header_token_count = 0 |
| 548 | |
| 549 | self.state = transition.next_state |
| 550 | |
| 551 | for event_type in transition.events: |
| 552 | if event_type == EventType.TOOL_CALL_START: |
| 553 | self.tool_index += 1 |
| 554 | event_value = ( |
| 555 | message_header |
| 556 | if previous_state == ParserState.MESSAGE_HEADER |
| 557 | and event_type == EventType.TEXT_CHUNK |
| 558 | else value |
| 559 | ) |
| 560 | if event_type == EventType.TEXT_CHUNK and not event_value: |
| 561 | continue |
| 562 | events.append( |
| 563 | SemanticEvent( |
| 564 | event_type, |
| 565 | value=event_value, |
| 566 | tool_index=self.tool_index, |
| 567 | token_count=( |
| 568 | message_header_token_count |
| 569 | if previous_state == ParserState.MESSAGE_HEADER |
| 570 | and event_type == EventType.TEXT_CHUNK |
| 571 | else token_count |
| 572 | ), |
| 573 | ) |
| 574 | ) |
| 575 | |
| 576 | if self.state == ParserState.TOOL_ARGS: |
| 577 | self._args_brace_depth = 0 |
| 578 | self._args_in_string = False |
| 579 | self._args_escape_next = False |
| 580 | self._args_safe_end = 0 |
| 581 | if self.config.tool_call_body_array: |
| 582 | self._reset_array_state() |
| 583 | |
| 584 | return events |
| 585 | |
| 586 | def _feed_args_text(self, text: str) -> list[SemanticEvent]: |
| 587 | """Feed text into the JSON argument streaming buffer. |
| 588 | |
| 589 | Streams argument characters incrementally while holding back |
| 590 | closing braces/brackets that might change as more input arrives. |
| 591 | """ |
| 592 | events: list[SemanticEvent] = [] |
| 593 | for ch in text: |
| 594 | result = self._feed_args_char(ch) |
| 595 | events.extend(result) |
| 596 | return events |
| 597 | |
| 598 | def _feed_args_char(self, ch: str) -> list[SemanticEvent]: |
| 599 | self._args_buffer += ch |
| 600 | |
| 601 | if self._args_escape_next: |
| 602 | self._args_escape_next = False |
| 603 | self._args_safe_end = len(self._args_buffer) |
| 604 | return self._flush_safe_args() |
| 605 | |
| 606 | if self._args_in_string: |
| 607 | if ch == "\\": |
| 608 | self._args_escape_next = True |
| 609 | elif ch == '"': |
| 610 | self._args_in_string = False |
| 611 | self._args_safe_end = len(self._args_buffer) |
| 612 | return self._flush_safe_args() |
| 613 | |
| 614 | if ch == '"': |
| 615 | self._args_in_string = True |
| 616 | self._args_safe_end = len(self._args_buffer) |
| 617 | return self._flush_safe_args() |
| 618 | |
| 619 | if ch in ("{", "["): |
| 620 | self._args_brace_depth += 1 |
| 621 | self._args_safe_end = len(self._args_buffer) |
| 622 | return self._flush_safe_args() |
| 623 | |
| 624 | if ch in ("}", "]"): |
| 625 | if self._args_brace_depth > 0: |
| 626 | self._args_brace_depth -= 1 |
| 627 | if self._args_brace_depth == 0: |
| 628 | return [] |
| 629 | self._args_safe_end = len(self._args_buffer) |
| 630 | return self._flush_safe_args() |
| 631 | |
| 632 | self._args_safe_end = len(self._args_buffer) |
| 633 | return self._flush_safe_args() |
| 634 | |
| 635 | def _flush_safe_args(self) -> list[SemanticEvent]: |
| 636 | """Emit buffered argument characters up to the safe-end watermark. |
| 637 | |
| 638 | Top-level closing braces are held back (safe_end not advanced) |
| 639 | until confirmed safe by a subsequent character or finish(). |
| 640 | """ |
| 641 | if self._args_safe_end == 0: |
| 642 | return [] |
| 643 | to_emit = self._args_buffer[: self._args_safe_end] |
| 644 | self._args_buffer = self._args_buffer[self._args_safe_end :] |
| 645 | self._args_safe_end = 0 |
| 646 | return [ |
| 647 | SemanticEvent( |
| 648 | EventType.ARG_VALUE_CHUNK, |
| 649 | value=to_emit, |
| 650 | tool_index=self.tool_index, |
| 651 | ) |
| 652 | ] |
| 653 | |
| 654 | # ── JSON-array tool body (tool_call_body_array) ───────────────────── |
| 655 | |
| 656 | def _feed_array_text(self, text: str) -> list[SemanticEvent]: |
| 657 | """Split a JSON-array tool body into one tool call per element. |
| 658 | |
| 659 | Granite emits ``[{"name":..,"arguments":{..}}, ...]`` after the tool |
| 660 | marker. Element interiors are streamed through the string/brace-aware |
| 661 | :meth:`_feed_args_char`, so each element behaves exactly like a |
| 662 | single-object tool body; the array's ``]`` ends the region and any |
| 663 | trailing text becomes content. |
| 664 | """ |
| 665 | events: list[SemanticEvent] = [] |
| 666 | for i, ch in enumerate(text): |
| 667 | if self.state != ParserState.TOOL_ARGS: |
| 668 | # The array closed mid-chunk; the rest of the text is content. |
| 669 | events.extend(self._emit_for_state(text[i:])) |
| 670 | break |
| 671 | events.extend(self._feed_array_char(ch)) |
| 672 | return events |
| 673 | |
| 674 | def _feed_array_char(self, ch: str) -> list[SemanticEvent]: |
| 675 | if self._array_in_element: |
| 676 | events = self._feed_args_char(ch) |
| 677 | # Element object closed: top-level '}' with all braces balanced. |
| 678 | if ( |
| 679 | ch == "}" |
| 680 | and not self._args_in_string |
| 681 | and not self._args_escape_next |
| 682 | and self._args_brace_depth == 0 |
| 683 | ): |
| 684 | events.extend(self._end_array_element()) |
| 685 | return events |
| 686 | |
| 687 | # Between elements only structural array characters are expected. |
| 688 | if ch in " \t\r\n": |
| 689 | return [] |
| 690 | if not self._array_started: |
| 691 | if ch == "[": |
| 692 | self._array_started = True |
| 693 | return [] |
| 694 | return self._bail_array_to_content(ch) |
| 695 | if ch == ",": |
| 696 | return [] |
| 697 | if ch == "{": |
| 698 | return self._start_array_element(ch) |
| 699 | if ch == "]": |
| 700 | self.state = ParserState.CONTENT |
| 701 | self._reset_array_state() |
| 702 | return [] |
| 703 | return self._bail_array_to_content(ch) |
| 704 | |
| 705 | def _start_array_element(self, ch: str) -> list[SemanticEvent]: |
| 706 | self.tool_index += 1 |
| 707 | self._reset_args_state() |
| 708 | self._array_in_element = True |
| 709 | events = [SemanticEvent(EventType.TOOL_CALL_START, tool_index=self.tool_index)] |
| 710 | events.extend(self._feed_args_char(ch)) |
| 711 | return events |
| 712 | |
| 713 | def _end_array_element(self) -> list[SemanticEvent]: |
| 714 | self._array_in_element = False |
| 715 | events: list[SemanticEvent] = [] |
| 716 | # Flush the held closing '}' so the slot body is the full element. |
| 717 | if self._args_buffer: |
| 718 | events.append( |
| 719 | SemanticEvent( |
| 720 | EventType.ARG_VALUE_CHUNK, |
| 721 | value=self._args_buffer, |
| 722 | tool_index=self.tool_index, |
| 723 | ) |
| 724 | ) |
| 725 | self._args_buffer = "" |
| 726 | self._args_safe_end = 0 |
| 727 | events.append( |
| 728 | SemanticEvent(EventType.TOOL_CALL_END, tool_index=self.tool_index) |
| 729 | ) |
| 730 | return events |
| 731 | |
| 732 | def _bail_array_to_content(self, ch: str) -> list[SemanticEvent]: |
| 733 | # Marker not followed by a JSON array (malformed) — revert to content |
| 734 | # so the remaining tokens are emitted as plain text, no tool call. |
| 735 | self.state = ParserState.CONTENT |
| 736 | self._reset_array_state() |
| 737 | return self._emit_for_state(ch) |