Incorrect speaker extracted in commentary and outlook section of dialogue

[pranshu-raj-211]

Cases where this is observed:

1. indusind bank
2. ambuja cements

Speaker is part of a sentence, dialogue is speech following that sentence.
Also noticed - Part of text by a speaker not extracted (page 4 of ambuja cements, jumped from page 3, incorrect speaker observed on page 5)

[pranshu-raj-211]

Changes made in PR #10 added this issue for apollo hospitals case.

Probable cause: Extraction of speech is done by a regex based search, which searches for a pattern like .:

Suggested fix: Verify if the string before colon is a name or not (easy task for non apollo like formats).

Need documentation for all types of document formats encountered, will help if issues arise later with inclusion of more documents.

[pranshu-raj-211]

Suggested fix:
Modify extract management agent, and the parser.extract_management_team function to use 2 contexts.

For example, modify extract_management.py line 40 as:

def process(page_text:str, speakers:list[str]):
    if page_text != "":
        context = CONTEXT   # original context - extracts the names and designations from the whole page text.
    else:
        context = alternative_context   # checks if the list of names given is names or some random sentence, returns only names
    messages = [
            {"role": "system", "content": context},
            {"role": "user", "content": page_text},
        ]

    response = get_groq_response(messages=messages, model=MODEL_NAME)

    return response

[pranshu-raj-211]

Just modify speaker regex, instead of being in the init method, declare it in the parse_conference_call method, with dynamic pattern that looks like :