A low-level Python-based question-answering system without using langchain, llama-index, etc. that uses CoT and ReAct (Reasoning and Acting) to provide accurate answers from documents. The system supports multiple document formats and employs an agent-based approach for intelligent information retrieval and response generation.

• ReAct-based agent for intelligent question answering
• CoT prompting

1. Clone the repository:

git clone https://github.com/yourusername/document-qa-agent.git
cd document-qa-agent

2. Install the required dependencies:

pip install -r requirements.txt

3. Store the OpenAPI Chat Model keys in the .env

from utils.document import Document
from agentqa import ReActDocumentQA

# Initialize with a document
document_name = 'example.pdf'
document_type = 'PDF'

# Create document object
doc_obj = Document(doc_name=document_name, type=document_type)

# Initialize the QA agent
agent = ReActDocumentQA(doc_obj.document, index_name='example')

# Ask a question
question = "What is the main topic?"
answer = agent.process_question(question)

You can process multiple questions at once using the run_app function:

questions = [
    "What is the main topic?",
    "Who are the key participants?"
]

result = run_app('document.pdf', 'PDF', questions)
print(result)

• PDF (.pdf)
• Text files (.txt)

# Example with a text file
document_name = 'lily_story.txt'
document_type = 'TXT'
questions = [
    "What Lily did one morning?",
    "What did Lily bring to the pond and why?"
]

result = run_app(document_name, document_type, questions)
print(result)

# Example with a PDF file
document_name = 'story.pdf'
document_type = 'PDF'
questions = [
        "Where did Clara live?",
        "What Clara and Leo did together?",
        "Tell me about adventure of Clara and Leo."
      ] 
result = run_app(document_name, document_type, questions)

The system returns results in JSON format:

{
    "questions": ["Question 1", "Question 2"],
    "answers": ["Answer 1", "Answer 2"]
}

1. Performance Dependencies:

   • The agent's performance is heavily dependent on prompt engineering and the language models used
   • Results can be improved significantly with better prompt engineering and more advanced language models

2. Document Processing Limitations:

   • The document searcher's performance varies based on document complexity
   • Current implementation works best with simple PDF documents where paragraph segmentation is straightforward
   • Document indexing should be optimized for better retrieval performance

1. Document Processing:

   • Implement advanced document segmentation techniques
   • Add support for more document formats
   • Develop better indexing mechanisms for complex documents
   • Optimize paragraph extraction and processing

2. Agent Enhancement:

   • Optimize prompt engineering for better question understanding and answer generation
   • Upgrade to more advanced language models
   • Implement better context management for follow-up questions
   • Add support for multi-document querying

3. Performance Optimization:

   • Implement caching mechanisms for frequently accessed documents
   • Optimize document search algorithms
   • Add parallel processing for batch questions
   • Implement better error handling and recovery mechanisms

4. User Interface:

   • Add a web interface for easier interaction
   • Implement real-time processing feedback
   • Add visualization for document segmentation
   • Develop better result formatting options