Code Flow and Library Choices

This section explains the flow of the code and the rationale behind using different libraries.

Core Logic

• Document Loading:

  • The data_ingestion() function loads documents from the DATA_DIR.

  • If the directory is empty, it copies the pre-loaded document.

  • Using SimpleDirectoryReader from LlamaIndex makes it easy to read files from a directory.

• Index Creation:

  • The VectorStoreIndex.from_documents() function creates a vector store index from the loaded documents using LlamaIndex.

  • This function takes the llm (LLM model) and embed_model (embedding model) as parameters, allowing for the use of specific models.

• Index Storage:

  • The index is saved to a persistent directory (PERSIST_DIR) using the StorageContext from LlamaIndex. This enables the index to be loaded across sessions.

• Query Handling:

  • The handle_query() function loads the index from storage or creates a new one if it doesn’t exist.

  • It then checks for greetings or feature requests.

  • For document-based questions, the query_engine from the index is used to get a response.

  • Cosine similarity is calculated between the query embedding and the response embedding to ensure the answer is contextually relevant.

Library Choices

• LlamaIndex: Chosen for its specialized capabilities in document-based Q&A, including indexing, storage, and retrieval.

• Streamlit: Used for building the user interface, providing a user-friendly chatbot experience.

• Hugging Face: Used for both the LLM (meta-llama/Meta-Llama-3-8B-Instruct) and the embedding model (BAAI/bge-small-en-v1.5).

• Scikit-learn: Used for calculating cosine similarity, which is used to determine the relevance of the answers.

This code flow highlights the key components and libraries used to create a robust and efficient document-based question-answering chatbot.