utils/text_processing.py

# Utility functions for GAIA Agent Evaluator
from langchain.docstore.document import Document
from langchain.text_splitter import RecursiveCharacterTextSplitter
from config import GAIA_KNOWLEDGE, ANSWER_PREFIXES_TO_REMOVE, LLM_RESPONSE_MARKERS, LLM_END_MARKERS

def create_knowledge_documents():
    """Create knowledge base documents from GAIA_KNOWLEDGE."""
    text_splitter = RecursiveCharacterTextSplitter(
        chunk_size=500,
        chunk_overlap=50,
        separators=["\n\n", "\n", ".", "!", "?", ",", " ", ""]
    )
    knowledge_chunks = text_splitter.split_text(GAIA_KNOWLEDGE)
    return [Document(page_content=chunk) for chunk in knowledge_chunks]

def clean_llm_response(response, prompt):
    """Clean up the LLM response to extract the answer."""
    # Remove the prompt from the beginning if it's included
    if response.startswith(prompt):
        response = response[len(prompt):]
    
    # Try to find where the model's actual answer begins
    for marker in LLM_RESPONSE_MARKERS:
        if marker.lower() in response.lower():
            parts = response.lower().split(marker.lower(), 1)
            if len(parts) > 1:
                response = parts[1].strip()
    
    # Remove any closing tags if they exist
    for marker in LLM_END_MARKERS:
        if marker.lower() in response.lower():
            response = response.lower().split(marker.lower())[0].strip()
    
    return response.strip()

def extract_final_answer(answer):
    """Extract and clean the final answer for exact matching."""
    clean_answer = answer.strip()
    
    # Remove prefixes (case insensitive)
    for prefix in ANSWER_PREFIXES_TO_REMOVE:
        if clean_answer.lower().startswith(prefix.lower()):
            clean_answer = clean_answer[len(prefix):].strip()
    
    # Remove quotes if the entire answer is quoted
    if clean_answer.startswith('"') and clean_answer.endswith('"'):
        clean_answer = clean_answer[1:-1]
    elif clean_answer.startswith("'") and clean_answer.endswith("'"):
        clean_answer = clean_answer[1:-1]
    
    # Remove trailing periods if they seem extraneous
    if clean_answer.endswith('.') and not clean_answer.replace('.', '').isdigit():
        # Don't remove decimal points from numbers
        if not (clean_answer.count('.') == 1 and clean_answer.replace('.', '').isdigit()):
            clean_answer = clean_answer[:-1]
    
    # Clean up extra whitespace
    clean_answer = ' '.join(clean_answer.split())
    
    return clean_answer

def format_prompt(question, context=""):
    """Format the question into a proper prompt for the LLM."""
    if context:
        return f"""You are a precise AI assistant that answers questions using available information. Your answer will be evaluated with exact string matching, so provide only the specific answer requested without additional text.

Context Information:
{context}

Question: {question}

Critical Instructions:
- Provide ONLY the exact answer requested, nothing else
- Do not include phrases like "The answer is", "Final answer", or "Based on the context"
- For numerical answers, use the exact format requested (integers, decimals, etc.)
- For lists, use the exact formatting specified in the question (commas, spaces, etc.)
- For names, use proper capitalization as would appear in official sources
- Be concise and precise - extra words will cause evaluation failure
- If the question asks for multiple items, provide them in the exact format requested

Direct Answer:"""
    else:
        return f"""You are a precise AI assistant that answers questions accurately. Your answer will be evaluated with exact string matching, so provide only the specific answer requested without additional text.

Question: {question}

Critical Instructions:
- Provide ONLY the exact answer requested, nothing else
- Do not include phrases like "The answer is", "Final answer", or explanations
- For numerical answers, use the exact format that would be expected
- For lists, use appropriate formatting (commas, spaces, etc.)
- For names, use proper capitalization
- Be concise and precise - extra words will cause evaluation failure
- Answer based on your knowledge and reasoning

Direct Answer:"""