DataBase Question Answering (DBQA)

DBQA is an application that leverages Retrieval-Augmented Generation (RAG) to provide answers to questions that are related to local documents.
It supports a wide array of document formats, such as txt, html, word, pdf, epub, and more.

Dependency Installation

For Ubuntu

apt-get install python-dev-is-python3 libxml2-dev libxslt1-dev antiword unrtf poppler-utils pstotext
pip install -r requirements.txt

For Archlinux

pacman -S python libxml2 libxslt antiword unrtf poppler pstotext
pip install -r requirements.txt

Usage

Note: The local TAIDE model is expected located at /var/models/llama2-7b-chat-b5.0.0

1. Construct the vector database from local documents

   python construct_vector_db.py /path/to/documents /path/to/output/directory

2. Ask question about your documents

   python dbqa.py /path/to/database/directory "Your question"

Development

Architecture

flowchart LR

subgraph "Retrival-Augmented Generation (RAG)"
    direction LR
    *A(Start) --> AA[/Query/]
    AB[(Vector DB)]
    AC[/Prompt template/]
    AA ---> AD[Retriever]
    AB ---> AD
    AA ---> AE[Reranker]
    AD -- Documents --> AE
    AC --> AF[Formater]
    AA ---> AF
    AE -- Related documents --> AF
    AF -- Prompt --> AG["Generator (LLM)"]
    AG --> AH[/Answer/]
end


subgraph Database Construction
    direction LR
    *B(Start) --> BA[/Raw documents/]
    BA --> BB[Text extractor]
    BB -- Docuemtnes --> BC[Splitter]
    BC -- Chunks --> BD[Embedding model] 
    BD -- Sentence\nEmbedding --> BE[(Vector DB)]
    BC -- Chunks --> BE 
end

Loading

Class Documentation

• ChatTuple: Grouped chat record for rendering prompt
• DocumentStore: Encapsulation of the vector database and the embedding model
• TaideChatModel: LangChain-compatible ChatModel for TAIDE
• TextractLoader: Universal file text extractor
• ParallelSplitter: Splitter with multiprocessing