examples/voice_agent/server/parsers/nemotron_toolcall_parser_streaming.py

# Copyright (c) 2025, NVIDIA CORPORATION.  All rights reserved.
#
# Licensed under the Apache License, Version 2.0 (the "License");
# you may not use this file except in compliance with the License.
# You may obtain a copy of the License at
#
#     http://www.apache.org/licenses/LICENSE-2.0
#
# Unless required by applicable law or agreed to in writing, software
# distributed under the License is distributed on an "AS IS" BASIS,
# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
# See the License for the specific language governing permissions and
# limitations under the License.

# Code adapted from https://huggingface.co/nvidia/NVIDIA-Nemotron-Nano-9B-v2/resolve/main/nemotron_toolcall_parser_streaming.py

import json
from collections.abc import Sequence
from random import choices
from string import ascii_letters, digits
from typing import Optional, Union

import partial_json_parser
import regex as re
from partial_json_parser.core.options import Allow
from pydantic import Field
from vllm.entrypoints.openai.protocol import (
    ChatCompletionRequest,
    DeltaFunctionCall,
    DeltaMessage,
    DeltaToolCall,
    ExtractedToolCallInformation,
    FunctionCall,
    ToolCall,
)
from vllm.logger import init_logger
from vllm.tokenizers.mistral import MistralTokenizer
from vllm.tool_parsers.abstract_tool_parser import ToolParser, ToolParserManager
from vllm.transformers_utils.tokenizer import AnyTokenizer

logger = init_logger(__name__)

ALPHANUMERIC = ascii_letters + digits


class NemotronToolCall(ToolCall):
    id: str = Field(default_factory=lambda: NemotronToolCall.generate_random_id())

    @staticmethod
    def generate_random_id():
        return "".join(choices(ALPHANUMERIC, k=9))

    @staticmethod
    def is_valid_id(id: str) -> bool:
        return id.isalnum() and len(id) == 9


def _is_fn_name_regex_support(model_tokenizer: AnyTokenizer) -> bool:
    return isinstance(model_tokenizer, MistralTokenizer) and model_tokenizer.version >= 11


@ToolParserManager.register_module("nemotron_json")
class NemotronToolParser(ToolParser):
    """
    Streaming tool call parser specifically designed for the Nemotron-Nano-V2 model.

    This parser functions as an active reconstruction engine, managing the realtime
    transition from text generation to structured tool execution. Its primary responsibilities
    during token streaming include:

    - Interception: Detects and consumes the `<TOOLCALL>` control tokens to switch parsing modes.
    - Buffering: Manages a lookahead buffer to prevent ambiguous partial tags (like `<TOO`)
                 from leaking to the user.
    - Restoration: Utilizes `partial_json_parser` to reconstruct valid objects from incomplete
                   JSON fragments.
    - Differentiation: Computes the precise "delta" between the current and previous JSON states
                       to ensure monotonic streaming.
    - Sanitization: Strips premature auto-completed closing characters (e.g., `}`)
                    to prevent malformed updates.

    Configuration:
        Activate this parser in the vLLM server by setting the following mandatory arguments:
        - `--enable-auto-tool-choice`
        - `--tool-call-parser nemotron_json`
    """

    def __init__(self, tokenizer: AnyTokenizer):
        super().__init__(tokenizer)
        # initialize properties used for state when parsing tool calls in
        # streaming mode
        self.prev_tool_call_arr: list[dict] = []
        self.current_tool_id: int = -1
        self.current_tool_name_sent: bool = False
        self.streamed_args_for_tool: list[str] = []  # map what has been streamed for each tool so far to a list
        self.tool_args_emitted: list[bool] = []
        self.bot_token = "<TOOLCALL>"
        self.bot_token_id = self.vocab.get(self.bot_token)
        logger.info(f"Nemotron Tool Parser: bot_token: {self.bot_token}, bot_token_id: {self.bot_token_id}")
        self.tool_call_regex = re.compile(r"\[{.*}\]", re.DOTALL)
        if _is_fn_name_regex_support(self.model_tokenizer):
            self.fn_name_regex = re.compile(r'([a-zA-Z0-9_-]+)(\{[\s\S]*?\})(?=\s*$|,|\s)', re.DOTALL)
        else:
            self.fn_name_regex = None

        # Buffer for partial tag sequences to disambiguate between normal content and
        # a forthcoming <TOOLCALL> or </TOOLCALL> tag in streaming.
        self._pending_tag_buffer: str = ""

    def _reset_state(self) -> None:
        """
        Reset the parser state for a new request.
        This is used to prevent state corruption across multiple sequential requests.
        """
        self.prev_tool_call_arr: list[dict] = []
        self.current_tool_id: int = -1
        self.current_tool_name_sent: bool = False
        self.streamed_args_for_tool: list[str] = []
        self.tool_args_emitted: list[bool] = []
        self._pending_tag_buffer: str = ""

    @staticmethod
    def _strip_trailing_auto_closers(chunk: str) -> str:
        """
        Remove parser auto-completed closing braces/brackets plus trailing whitespace.
        These should be flushed only when a tool call completes to avoid duplicate
        argument fragments.

        Args:
            chunk (str):
                The chunk of text to strip.
        Return:
            (str): The chunk of text with trailing auto-completed closing braces/brackets
                   plus trailing whitespace removed.
        """
        idx = len(chunk)
        while idx > 0 and chunk[idx - 1] in " \t\r\n}]":
            idx -= 1
        # Remove trailing non-escaped double quotes (partial JSON auto-closes strings)
        while idx > 0 and chunk[idx - 1] == '"':
            # keep escaped quotes (\"), only strip bare ones
            if idx - 2 >= 0 and chunk[idx - 2] == '\\':
                break
            idx -= 1
        return chunk[:idx]

    @staticmethod
    def _common_prefix_len(left: str, right: str) -> int:
        """
        Calculate the length of the longest initial substring shared by two strings.

        This utility is used to determine how much of the tool arguments have already
        been streamed to the client, allowing the system to send only the new 'delta'.

        Args:
            left (str): The first string to compare (typically the full current arguments).
            right (str): The second string to compare (typically the previously streamed arguments).

        Returns:
            int: The count of identical characters starting from index 0.
                 Returns 0 if the strings share no common prefix.
        """
        max_len = min(len(left), len(right))
        idx = 0
        while idx < max_len and left[idx] == right[idx]:
            idx += 1
        return idx

    def _compute_arguments_delta(self, cur_arguments_json: str, end_of_call: bool) -> str:
        """
        Determine the incremental suffix to stream for the current tool call.
        Ensures we only emit monotonic chunks by trimming our tracked prefix to
        the longest common prefix with the latest JSON snapshot.

        Args:
            cur_arguments_json (str):
                The current arguments JSON in string format.
            end_of_call (bool):
                Whether the current tool call is the last one in the array.

        Return:
            (str): The incremental suffix to stream for the current tool call.
        """
        tool_idx = self.current_tool_id
        if tool_idx < 0 or tool_idx >= len(self.streamed_args_for_tool):
            if tool_idx < 0:
                logger.debug(f"current_tool_id is negative ({tool_idx}), no tool designated yet")
            else:
                logger.warning(
                    f"tool_idx ({tool_idx}) is out of bounds for streamed_args_for_tool "
                    f"(length: {len(self.streamed_args_for_tool)})"
                )
            return ""

        streamed_prefix = self.streamed_args_for_tool[tool_idx]
        had_any = self.tool_args_emitted[tool_idx] if tool_idx < len(self.tool_args_emitted) else False

        lcp_len = self._common_prefix_len(cur_arguments_json, streamed_prefix)
        if lcp_len != len(streamed_prefix):
            streamed_prefix = streamed_prefix[:lcp_len]
            self.streamed_args_for_tool[tool_idx] = streamed_prefix

        if (
            not had_any
            and not end_of_call
            and lcp_len == 0
            and cur_arguments_json.endswith('": ""}')
            and '": ""' in cur_arguments_json
        ):
            closing_pos = cur_arguments_json.rfind('": ""}')
            if closing_pos != -1:
                arguments_delta = cur_arguments_json[: closing_pos + 4]
            else:
                arguments_delta = cur_arguments_json
        else:
            arguments_delta = cur_arguments_json[lcp_len:]

        if not arguments_delta:
            return ""

        if not end_of_call:
            arguments_delta = self._strip_trailing_auto_closers(arguments_delta)

        if not had_any and not end_of_call and arguments_delta and arguments_delta.endswith('}'):
            arguments_delta = arguments_delta[:-1]
            if arguments_delta.endswith('"'):
                arguments_delta = arguments_delta[:-1]

        return arguments_delta

    def _visible_delta_outside_tool(
        self, delta_text: str, start_token: Optional[str], end_token: Optional[str]
    ) -> str:
        """
        Filters incoming streaming text to hide incomplete or complete tool call tags.

        This method acts as a buffer for the streaming response. It consumes and holds
        characters that resemble the start of `start_token` or `end_token` (e.g., "<", "<T", "<TOO").

        - If the buffer eventually matches the full token exactly (e.g., "<TOOLCALL>"),
          the buffer is discarded (suppressed).
        - If the buffer diverges from the expected tokens (e.g., user types "<Think>"),
          the buffered text is released (flushed) alongside the current character.
        - Regular text that does not start with "<" passes through immediately.

        Args:
            delta_text (str):
                The new chunk of text generated by the model in this streaming step.
            start_token (Optional[str]):
                The opening tag to suppress (e.g., "<TOOLCALL>"). If None, no start tag is tracked.
            end_token (Optional[str]):
                The closing tag to suppress (e.g., "</TOOLCALL>"). If None, no end tag is tracked.

        Returns:
            str: The portion of `delta_text` (plus any previously buffered ambiguous characters)
            that has been confirmed as *not* being part of a tool call tag.
        """
        if not delta_text:
            return delta_text

        visible: list[str] = []
        for ch in delta_text:
            if self._pending_tag_buffer or ch == '<':
                self._pending_tag_buffer += ch

                if start_token and start_token.startswith(self._pending_tag_buffer):
                    if self._pending_tag_buffer == start_token:
                        self._pending_tag_buffer = ""
                    continue

                if end_token and end_token.startswith(self._pending_tag_buffer):
                    if self._pending_tag_buffer == end_token:
                        self._pending_tag_buffer = ""
                    continue

                # Not a tool tag; flush buffered characters as normal content.
                visible.append(self._pending_tag_buffer)
                self._pending_tag_buffer = ""
            else:
                visible.append(ch)

        return "".join(visible)

    def adjust_request(self, request: ChatCompletionRequest) -> ChatCompletionRequest:
        if not isinstance(self.model_tokenizer, MistralTokenizer) and request.tools and request.tool_choice != 'none':
            # Do not skip special tokens when using chat template
            # with Mistral parser as TOOL_CALL token is needed
            # for tool detection.
            # Note: we don't want skip_special_tokens=False
            # with MistralTokenizer as it is incompatible
            request.skip_special_tokens = False
        return request

    def extract_tool_calls(
        self,
        model_output: str,
        request: ChatCompletionRequest,
    ) -> ExtractedToolCallInformation:
        """
        Parses a complete (non-streaming) model response to extract tool execution instructions.

        This method attempts to convert the raw text output from the model into structured
        `NemotronToolCall` objects. It employs a robust two-stage parsing strategy:
        - Direct JSON Parsing: First attempts to parse the content following the
           `<TOOLCALL>` token as valid JSON.
        - Regex Fallback: If direct parsing fails (e.g., due to extra text or noise),
           it uses a regular expression to locate and extract the specific JSON array pattern.

        Args:
            model_output (str): The full text generated by the model.
            request (ChatCompletionRequest): The original request object (used for context if needed).

        Returns:
            ExtractedToolCallInformation: An object containing the parsed list of tool calls
            and any preceding text content. If parsing fails entirely, it returns the raw
            content as a standard text message.
        """
        # Reset state for each new non-streaming request
        self._reset_state()

        # case -- if a tool call token is not present, return a text response
        if self.bot_token not in model_output:
            return ExtractedToolCallInformation(tools_called=False, tool_calls=[], content=model_output)

        # first remove the BOT token
        tool_content = model_output.replace(self.bot_token, "").strip()

        try:
            # we first try to directly load the json as parsing very nested
            # jsons is difficult
            try:
                if self.fn_name_regex:
                    matches = self.fn_name_regex.findall(tool_content)

                    function_call_arr = []
                    for match in matches:
                        fn_name = match[0]
                        args = match[1]

                        # fn_name is encoded outside serialized json dump
                        # only arguments are serialized
                        function_call_arr.append({"name": fn_name, "arguments": json.loads(args)})
                else:
                    function_call_arr = json.loads(tool_content)
            except json.JSONDecodeError:
                # use a regex to find the part corresponding to the tool call.
                # NOTE: This use case should not happen if the model is trained
                # correctly. It's an easy possible fix so it's included, but
                # can be brittle for very complex / highly nested tool calls
                matches = self.tool_call_regex.findall(tool_content)
                if not matches:
                    raise ValueError(f"No tool call pattern found in: {tool_content[:100]} ...")
                raw_tool_call = matches[0]
                function_call_arr = json.loads(raw_tool_call)

            # Tool Call
            tool_calls: list[NemotronToolCall] = [
                NemotronToolCall(
                    type="function",
                    function=FunctionCall(
                        name=raw_function_call["name"],
                        # function call args are JSON but as a string
                        arguments=json.dumps(raw_function_call["arguments"], ensure_ascii=False),
                    ),
                )
                for raw_function_call in function_call_arr
            ]

            # get any content before  the tool call
            content = model_output.split(self.bot_token)[0]
            return ExtractedToolCallInformation(
                tools_called=True, tool_calls=tool_calls, content=content if len(content) > 0 else None
            )

        except Exception:
            logger.exception("Error in extracting tool call from response.")
            # return information to just treat the tool call as regular JSON
            return ExtractedToolCallInformation(tools_called=False, tool_calls=[], content=tool_content)

    def extract_tool_calls_streaming(
        self,
        previous_text: str,
        current_text: str,
        delta_text: str,
        previous_token_ids: Sequence[int],
        current_token_ids: Sequence[int],
        delta_token_ids: Sequence[int],
        request: ChatCompletionRequest,
    ) -> Union[DeltaMessage, None]:
        """
        Parses the raw text stream to identify and extract tool calls in real-time.
        This method monitors for the `<TOOLCALL>` trigger to switch parsing modes, buffers
        ambiguous tag prefixes to prevent leakage, and utilizes partial JSON parsing to
        compute and emit precise incremental updates (deltas) for tool arguments while
        suppressing auto-generated artifacts.

        Args:
            previous_text (str): (Placeholder) The generated text prior to the current step.
            current_text (str): The total generated text including the new token.
            delta_text (str): The specific text chunk generated in this step.
            previous_token_ids (Sequence[int]): (Placeholder) Token IDs for previous text.
            current_token_ids (Sequence[int]): (Placeholder) Token IDs for current text.
            delta_token_ids (Sequence[int]): (Placeholder) Token IDs for the delta.
            request (ChatCompletionRequest): (Placeholder) The original client request object.

        Returns:
            Union[DeltaMessage, None]: A `DeltaMessage` containing visible content or
            tool call updates, or `None` if the output is currently buffered or unchanged.
        """
        # Reset state at the start of a new streaming request
        # Detect new request: if we have stale state but previous_text indicates this is a fresh start
        if not previous_text and (
            self.current_tool_id != -1 or self.prev_tool_call_arr or self.streamed_args_for_tool
        ):
            logger.debug("Detected new streaming request, resetting parser state")
            self._reset_state()

        # if candidates tool call tokens are in the tokens generated so far, that
        # means we're parsing as tool calls now. Suppress streaming if we are
        # currently generating any prefix of the start or end tag.
        visible_delta_text = delta_text
        try:
            start_token = self.bot_token
            end_token = f"</{self.bot_token[1:]}" if self.bot_token.startswith('<') else None

            visible_delta_text = self._visible_delta_outside_tool(delta_text, start_token, end_token)
        except Exception:
            # Fallback to conservative checks in case of any issues
            if (
                current_text.endswith('<')
                or current_text.endswith('<T')
                or current_text.endswith('<TO')
                or current_text.endswith('<TOOL')
                or current_text.endswith('<TOOLCALL')
            ):
                return None

        # if the tool call token is not in the tokens generated so far, append
        # output to contents since it's not a tool
        if self.bot_token not in current_text:
            if visible_delta_text:
                return DeltaMessage(content=visible_delta_text)
            # still waiting on a potential tag, so emit nothing yet
            return None

        # bit mask flags for partial JSON parsing. If the name hasn't been
        # sent yet, don't allow sending
        # an incomplete string since OpenAI only ever (as far as I have
        # seen) allows sending the entire tool/ function name at once.
        flags = Allow.ALL if self.current_tool_name_sent else Allow.ALL & ~Allow.STR
        end_of_call: bool = False
        try:

            # replace BOT token with empty string, and convert single quotes
            # to double to allow parsing as JSON since mistral uses single
            # quotes instead of double for tool calls
            parsable_arr = current_text.split(self.bot_token)[-1]

            # Check if we're at the end of the tool call
            if '</TOOLCALL>' in parsable_arr:
                end_of_call = True
                parsable_arr = parsable_arr.split('</TOOLCALL>')[0]

            # tool calls are generated in an array, so do partial JSON
            # parsing on the entire array
            try:
                tool_call_arr: list[dict] = partial_json_parser.loads(parsable_arr, flags)
            except (partial_json_parser.core.exceptions.MalformedJSON, json.JSONDecodeError, ValueError):
                return None

            current_tool_call: dict = (
                tool_call_arr[self.current_tool_id]
                if len(tool_call_arr) > 0 and self.current_tool_id >= 0 and self.current_tool_id < len(tool_call_arr)
                else {}
            )

            # case -- if no tokens have been streamed for the tool, e.g.
            #   only the array brackets, stream nothing
            if len(tool_call_arr) == 0:
                return None

            # case: we are starting a new tool in the array
            #   -> array has > 0 length AND length has moved past cursor
            elif len(tool_call_arr) > 0 and len(tool_call_arr) > self.current_tool_id + 1:

                # if we're moving on to a new call, first make sure we
                # haven't missed anything in the previous one that was
                # auto-generated due to JSON completions, but wasn't
                # streamed to the client yet.
                if self.current_tool_id >= 0 and self.current_tool_id < len(self.streamed_args_for_tool):
                    diff: Union[str, None] = current_tool_call.get("arguments")

                    if diff:
                        diff = json.dumps(diff, ensure_ascii=False).replace(
                            self.streamed_args_for_tool[self.current_tool_id], ""
                        )
                        delta = DeltaMessage(
                            tool_calls=[
                                DeltaToolCall(
                                    index=self.current_tool_id,
                                    function=DeltaFunctionCall(arguments=diff).model_dump(exclude_none=True),
                                )
                            ]
                        )
                        self.streamed_args_for_tool[self.current_tool_id] += diff
                    else:
                        delta = None
                else:
                    delta = None
                # re-set stuff pertaining to progress in the current tool
                self.current_tool_id = len(tool_call_arr) - 1
                self.current_tool_name_sent = False
                self.streamed_args_for_tool.append("")
                self.tool_args_emitted.append(False)
                return delta

            # case: update an existing tool - this is handled below

            # if the current tool name hasn't been sent, send if available
            # - otherwise send nothing
            if not self.current_tool_name_sent:
                function_name = current_tool_call.get("name")
                if function_name:

                    delta = DeltaMessage(
                        tool_calls=[
                            DeltaToolCall(
                                index=self.current_tool_id,
                                type="function",
                                id=NemotronToolCall.generate_random_id(),
                                function=DeltaFunctionCall(name=function_name).model_dump(exclude_none=True),
                            )
                        ]
                    )
                    self.current_tool_name_sent = True
                else:
                    delta = None

            # now we know we're on the same tool call and we're streaming
            # arguments
            else:
                if self.current_tool_id < 0 or self.current_tool_id >= len(self.prev_tool_call_arr):
                    prev_arguments = None
                else:
                    prev_arguments = self.prev_tool_call_arr[self.current_tool_id].get("arguments")
                cur_arguments = current_tool_call.get("arguments")

                if not cur_arguments and not prev_arguments:
                    delta = None
                elif not cur_arguments and prev_arguments:
                    logger.error("INVARIANT - impossible to have arguments reset " "mid-arguments")
                    delta = None
                elif cur_arguments:
                    cur_arguments_json = json.dumps(cur_arguments, ensure_ascii=False)
                    arguments_delta = self._compute_arguments_delta(cur_arguments_json, end_of_call)
                    if arguments_delta:
                        delta = DeltaMessage(
                            tool_calls=[
                                DeltaToolCall(
                                    index=self.current_tool_id,
                                    function=DeltaFunctionCall(arguments=arguments_delta).model_dump(
                                        exclude_none=True
                                    ),
                                )
                            ]
                        )
                        if self.current_tool_id >= 0 and self.current_tool_id < len(self.streamed_args_for_tool):
                            self.streamed_args_for_tool[self.current_tool_id] += arguments_delta
                        else:
                            logger.warning(
                                f"current_tool_id ({self.current_tool_id}) is out of bounds for streamed_args_for_tool "
                                f"(length: {len(self.streamed_args_for_tool)})"
                            )
                        if self.current_tool_id >= 0 and self.current_tool_id < len(self.tool_args_emitted):
                            self.tool_args_emitted[self.current_tool_id] = True
                        else:
                            logger.warning(
                                f"current_tool_id ({self.current_tool_id}) is out of bounds for tool_args_emitted "
                                f"(length: {len(self.tool_args_emitted)})"
                            )
                    else:
                        # Do not flush final JSON here; let the serving layer
                        # compute a minimal remaining suffix on finish.
                        delta = None
                else:
                    # End-of-call or equal state; do not force a final flush here.
                    delta = None

            # check to see if the name is defined and has been sent. if so,
            # stream the name - otherwise keep waiting
            # finish by setting old and returning None as base case
            self.prev_tool_call_arr = tool_call_arr
            # If we've reached the end of a tool call, flush any remaining
            # suffix (including a final '}') that hasn't been streamed yet.
            if end_of_call and self.current_tool_id >= 0:
                try:
                    cur_arguments = current_tool_call.get("arguments")
                    if cur_arguments is not None:
                        cur_args_json = json.dumps(cur_arguments, ensure_ascii=False)
                        remaining_suffix = self._compute_arguments_delta(cur_args_json, end_of_call=True)

                        # Only send remaining suffix if it's non-empty and contains meaningful content
                        # (not just whitespace or single characters like closing braces)
                        if remaining_suffix and remaining_suffix.strip():
                            extra = DeltaToolCall(
                                index=self.current_tool_id,
                                function=DeltaFunctionCall(arguments=remaining_suffix).model_dump(exclude_none=True),
                            )
                            if delta is None:
                                delta = DeltaMessage(tool_calls=[extra])
                            else:
                                if getattr(delta, "tool_calls", None):
                                    delta.tool_calls.append(extra)
                                else:
                                    delta.tool_calls = [extra]
                            if self.current_tool_id >= 0 and self.current_tool_id < len(self.streamed_args_for_tool):
                                self.streamed_args_for_tool[self.current_tool_id] += remaining_suffix
                            else:
                                logger.warning(
                                    f"current_tool_id ({self.current_tool_id}) is out of bounds for streamed_args_for_tool "
                                    f"(length: {len(self.streamed_args_for_tool)})"
                                )
                            if self.current_tool_id >= 0 and self.current_tool_id < len(self.tool_args_emitted):
                                self.tool_args_emitted[self.current_tool_id] = True
                            else:
                                logger.warning(
                                    f"current_tool_id ({self.current_tool_id}) is out of bounds for tool_args_emitted "
                                    f"(length: {len(self.tool_args_emitted)})"
                                )
                except Exception as e:
                    # Failure to flush the remaining arguments suffix is non-fatal; log for debugging.
                    logger.warning(f"Error in flushing remaining suffix for tool call: {e}")

            return delta

        except Exception as e:
            logger.exception(f"Error trying to handle streaming tool call: {e}")
            return None