Contextual Retrieval with Llama-Index (Anthropic)

This repository provides an implementation of contextual retrieval, a novel approach that enhances the performance of retrieval systems by incorporating chunk-specific explanatory context. By prepending contextual information to each chunk before embedding and indexing, this method improves the relevance and accuracy of retrieved results.

Key Technologies

• Llama-Index: A powerful framework for building semantic search applications.
• Ollama: A local LLMs serving solution, using the gemma2:2b model.
• Streamlit: A Python framework for building interactive web applications.
• FastAPI: A high-performance API framework for building web applications.
• ChromaDB: A vector database for efficient storage and retrieval of high-dimensional embeddings.

Examples

Example 1:

Example 2:

Setup

Clone the GitHub repo

git clone https://github.com/RionDsilvaCS/contextual-retrieval-by-anthropic.git

cd contextual-retrieval-by-anthropic

Create Python env and run requirements.txt file

pip install -r requirements.txt

Create a directory data and add all the PDF's there

mkdir data

Create .env file and add below variables.

DATA_DIR="./data"
SAVE_DIR="./src/db"
VECTOR_DB_PATH="./src/db/cook_book_db_vectordb"
BM25_DB_PATH="./src/db/cook_book_db_bm25"
COLLECTION_NAME="add_collection_name"
API_URL="http://127.0.0.1:8000/rag-chat"

Run Python file create_save_db.py to create ChromaDB and BM25 databases

python create_save_db.py

Run the application

Begin with running Ollama in separate terminal

ollama serve

Run the python file app.py to boot up FastAPI server

python app.py

Run the python file main.py to start streamlit app

streamlit run main.py

Additional Information

• Contextual Embedding: The process of prepending chunk-specific explanatory context to each chunk before embedding.
• Contextual BM25: A modified version of BM25 that incorporates contextual information for improved relevance scoring.

---

Follow me

GitHub @RionDsilvaCS · Linkedin @Rion Dsilva · Twitter @Rion_Dsilva_CS