Skip to content

README.md

Latest commit

 

History

History
69 lines (51 loc) · 2.39 KB

README.md

File metadata and controls

69 lines (51 loc) · 2.39 KB
Raw

News RAG System (Retrieval-Augmented Generation)

This project implements a Retrieval-Augmented Generation (RAG) pipeline over a news dataset (information from BBC News). Given a user question, the system:

  1. retrieves the most relevant news items via similarity search, and
  2. injects those items into a prompt to help an LLM produce a grounded answer.

The notebook walks through dataset loading, retrieval, prompt construction, and an end-to-end LLM call.

What’s inside

Core components implemented in the notebook

  • query_news(indices): fetch documents from the dataset by index.
  • retrieve(query, top_k): returns indices of the top-k most similar documents (provided via utils).
  • get_relevant_data(query, top_k): runs retrieval + fetches the corresponding rows.
  • format_relevant_data(relevant_data): formats retrieved items as readable context (title, description, date, URL).
  • generate_final_prompt(query, top_k, use_rag, prompt): builds the final prompt with optional RAG context.
  • llm_call(query, ...): calls the LLM with the final prompt.

Utilities (utils.py)

  • Loads a SentenceTransformer embedding model: BAAI/bge-base-en-v1.5
  • Loads precomputed embeddings from embeddings.joblib
  • Computes cosine similarity to retrieve top-k most relevant articles
  • Provides a small ipywidgets UI to compare answers with/without RAG
  • Supports LLM calls via:
    • a proxy endpoint (if no API key is provided), or
    • Together API (if TOGETHER_API_KEY is set)

Dataset

The notebook expects a deduplicated CSV:

  • news_data_dedup.csv

Key fields used:

  • title
  • description
  • url
  • published_at

Note: The retrieval context is treated as “additional information” to enrich the answer, not the only source of truth.

Requirements

1) Required files (same folder as the notebook)

  • news_data_dedup.csv (loaded from ./news_data_dedup.csv)
  • embeddings.joblib (loaded from ./embeddings.joblib)
  • utils.py
  • News_RAG.ipynb
  • unittests.py (optional, only if you want to run tests)

2) Python

  • Python 3.9+ (recommended)

3) Environment variables

Required (for embeddings model)

utils.py loads the SentenceTransformer model from a local directory:

  • MODEL_PATH must point to a folder that contains: BAAI/bge-base-en-v1.5

Example:

export MODEL_PATH="/path/to/models"
# Model must exist at:
# $MODEL_PATH/BAAI/bge-base-en-v1.5