[Bug]: regex pattern does not handle nested tags in prompt (structured compression)

[Imatgay]

Describe the bug

The original regular expression using ([^<]+)<\llmlingua> would fail when the text inside llmlingua tags contained other tags (like <tag>...</tag>). I suggest to replace it with something like ((?:[^<]*(?:<(?!/llmlingua>)[^>]*>)?)*?)</llmlingua> .

Commit on my fork: 73baf3f

Steps to reproduce

Confronting pattern matching:

from llmlingua import PromptCompressor

llm_lingua = PromptCompressor(
    model_name="microsoft/llmlingua-2-xlm-roberta-large-meetingbank",
    use_llmlingua2=True, 
    device_map='cpu',
)

structured_prompt_repo_example ="""<llmlingua, compress=False>Speaker 4:</llmlingua><llmlingua, rate=0.4> Thank you. And can we do the functions for content? </llmlingua><llmlingua, compress=False>
Speaker 0:</llmlingua><llmlingua, rate=0.4> Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group </llmlingua><llmlingua, compress=False>
Speaker 4:</llmlingua><llmlingua, rate=0.6> We have a promotion and a second time as councilman served Councilman Ringa and customers and they have any comments.</llmlingua>"""

original_pattern = r"<llmlingua\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*>([^<]+)</llmlingua>"

matches = re.findall(original_pattern, structured_prompt_repo_example)
print(matches)

# output: [('', 'False', '', '', 'Speaker 4:'), ('0.4', '', '', '', ' Thank you. And can we do the functions for content? '), 
#          ('', 'False', '', '', '\nSpeaker 0:'), ('0.4', '', '', '', ' Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group '), 
#          ('', 'False', '', '', '\nSpeaker 4:'), ('0.6', '', '', '', ' We have a promotion and a second time as councilman served Councilman Ringa and customers and they have any comments.')]

structured_prompt_with_nested_tags = """<llmlingua, compress=False>Speaker 4:</llmlingua><llmlingua, rate=0.4> Thank you. And can we do the functions for content? </llmlingua><llmlingua, compress=False>
Speaker 0:</llmlingua><llmlingua, rate=0.4> Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group </llmlingua><llmlingua, compress=False>
Speaker 4:</llmlingua><llmlingua, rate=0.6> We have a promotion <tag> and a second time as councilman served Councilman Ringa  </tag> and customers and they have any comments.</llmlingua>"""

matches = re.findall(original_pattern, structured_prompt_with_nested_tags )
print(matches)

# output: [('', 'False', '', '', 'Speaker 4:'), ('0.4', '', '', '', ' Thank you. And can we do the functions for content? '), 
#          ('', 'False', '', '', '\nSpeaker 0:'), ('0.4', '', '', '', ' Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group '), 
#          ('', 'False', '', '', '\nSpeaker 4:')] 

new_pattern = r"<llmlingua\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*(?:,\s*rate\s*=\s*([\d\.]+))?\s*(?:,\s*compress\s*=\s*(True|False))?\s*>((?:[^<]*(?:<(?!/llmlingua>)[^>]*>)?)*?)</llmlingua>"
matches = re.findall(new_pattern, structured_prompt_with_nested_tags)
print(matches)

# output: [('', 'False', '', '', 'Speaker 4:'), ('0.4', '', '', '', ' Thank you. And can we do the functions for content? '), 
#          ('', 'False', '', '', '\nSpeaker 0:'), ('0.4', '', '', '', ' Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group '), 
#          ('', 'False', '', '', '\nSpeaker 4:'), ('0.6', '', '', '', ' We have a promotion <tag> and a second time as councilman served Councilman Ringa  </tag> and customers and they have any comments.')]

Confronting "compressions" with/without nested non-llmlingua-related tags:

compressed_prompt_repo_example = llm_lingua.structured_compress_prompt(structured_prompt_repo_example, instruction="", question="", rate=0.5)
print(compressed_prompt_repo_example['compressed_prompt'])
# output: Speaker 4: Thank you. And can we do the functions for content? 
#         Speaker 0: Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group 
#         Speaker 4: We have a promotion and a second time as councilman served Councilman Ringa and customers and they have any comments.


compressed_prompt_with_nested_tags = llm_lingua.structured_compress_prompt(structured_prompt_with_nested_tags, instruction="", question="", rate=0.5)
print(compressed_prompt_with_nested_tags['compressed_prompt'])
# output:  Speaker 4: Thank you. And can we do the functions for content? 
#          Speaker 0: Item 11 is a communication from Council on Price recommendation to increase appropriation in the general fund group 
#          Speaker 4:

Expected Behavior

The regular expression should handle nested tags within the prompt without disrupting the matching process, treating them as plain text content.

Logs

No response

Additional Information

No response

[iofu728]

Hi @Imatgay, thanks for your feedback. We'll try to fix it soon.