Skip to content

Agent Utils

Utility functions and helpers for agent implementations.

rllm.agents.utils

get_recent_assistant_user_messages 

get_recent_assistant_user_messages(chat_completions_messages)

Extracts the most recent assistant message and environment messages (user/tool) from a chat completions list.

Parameters:

Name Type Description Default
chat_completions_messages List[Dict]

List of message dictionaries from chat completions.

 required

Returns:

Type Description

Tuple[Dict, List[Dict]]: A tuple containing: - The most recent assistant message (or None if not found) - A list of environment messages (user/tool) that occurred after the last assistant message, in chronological order.
Source code in rllm/agents/utils.py 
 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
def get_recent_assistant_user_messages(chat_completions_messages):
    """
    Extracts the most recent assistant message and environment messages (user/tool) from a chat completions list.

    Args:
        chat_completions_messages (List[Dict]): List of message dictionaries from chat completions.

    Returns:
        Tuple[Dict, List[Dict]]: A tuple containing:
            - The most recent assistant message (or None if not found)
            - A list of environment messages (user/tool) that occurred after the last assistant message,
              in chronological order.
    """
    # Loop backwards to get the last assistant message and environment messages
    env_messages = []
    assistant_message = None
    seen_assistant_message = False
    for message in reversed(chat_completions_messages):
        role = message.get("role", None)
        if role == "assistant":
            if assistant_message:
                break
            seen_assistant_message = True
            assistant_message = message
        elif role in ["user", "tool"] and not seen_assistant_message:
            env_messages.append(message)
    # Reverse the env_messages to maintain chronological order
    env_messages = list(reversed(env_messages))

    return assistant_message, env_messages

convert_messages_to_tokens_and_masks 

convert_messages_to_tokens_and_masks(messages: list[dict[str, str]], tokenizer: PreTrainedTokenizerBase, parser: ChatTemplateParser, contains_first_msg=False, contains_generation_msg=False)

Converts multiple messages to tokens and masks. contains_first_msg flag and contains_generaiton_msg flag are used to indicate whether the conversation is for beginning or contains the generation. The first and last message is assumed to be the special message respectively

Parameters:

Name Type Description Default
messages List[Dict]

The messages to convert.

 required
tokenizer PreTrainedTokenizerBase

The tokenizer to use.

 required
contains_first_msg bool

Whether the first message is a special message.

 False
contains_generation_msg bool

Whether the last message is a special message.

 False

Returns:

Type Description

Tuple[List[int], List[int]]: A tuple containing all tokens and all masks.
Source code in rllm/agents/utils.py 
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
def convert_messages_to_tokens_and_masks(messages: list[dict[str, str]], tokenizer: PreTrainedTokenizerBase, parser: ChatTemplateParser, contains_first_msg=False, contains_generation_msg=False):
    """
    Converts multiple messages to tokens and masks.
    contains_first_msg flag and contains_generaiton_msg flag are used to indicate whether the conversation is for beginning or contains the generation.
    The first and last message is assumed to be the special message respectively

    Args:
        messages (List[Dict]): The messages to convert.
        tokenizer: The tokenizer to use.
        contains_first_msg (bool): Whether the first message is a special message.
        contains_generation_msg (bool): Whether the last message is a special message.

    Returns:
        Tuple[List[int], List[int]]: A tuple containing all tokens and all masks.
    """
    all_msg_tokens = []
    all_msg_masks = []

    def _convert_message_to_tokens_and_masks(msg, first_msg=False, generation_msg=False):
        msg_text = parser.parse([msg], add_generation_prompt=generation_msg, is_first_msg=first_msg)

        # Remove the assistant token since it is contained in previous message as generation prompt
        if msg["role"] == "assistant":
            assert msg_text.startswith(parser.assistant_token), f"Expected assistant token {parser.assistant_token} but got {msg_text}"
            msg_text = msg_text.replace(parser.assistant_token, "")

        msg_tokens = tokenizer.encode(msg_text, add_special_tokens=False)
        mask_value = 1 if msg["role"] == "assistant" else 0
        msg_mask = [mask_value] * len(msg_tokens)

        return msg_tokens, msg_mask

    for i, msg in enumerate(messages):
        msg_tokens, msg_mask = _convert_message_to_tokens_and_masks(msg, first_msg=(contains_first_msg and i == 0), generation_msg=(contains_generation_msg and i == len(messages) - 1))
        all_msg_tokens.extend(msg_tokens)
        all_msg_masks.extend(msg_mask)

    return all_msg_tokens, all_msg_masks