Document RAG Assistant

A command-line retrieval-augmented generation application that answers questions using a local document collection. It creates document embeddings with Sentence Transformers, stores them in ChromaDB, retrieves relevant passages, and sends grounded context to a Hugging Face-hosted language model.

Features

• Ingests .txt and .pdf documents
• Splits documents into overlapping chunks
• Generates normalized embeddings with all-MiniLM-L6-v2
• Persists vectors and source metadata in ChromaDB
• Retrieves the three most relevant chunks for each question
• Constrains responses to retrieved context
• Displays the source documents used for an answer

Technology

• Python
• Hugging Face Inference API
• Sentence Transformers
• ChromaDB
• PyPDF

Setup

Create and activate a virtual environment, then install the dependencies:

python -m venv .venv
python -m pip install -r requirements.txt

Place .txt or .pdf files in documents_project4/, then build the local vector index:

python ingest.py

Start the assistant:

python main.py

The application accepts a Hugging Face token securely at runtime. Alternatively, define HF_TOKEN or HUGGINGFACEHUB_API_TOKEN in your environment. Never commit a token to the repository.

The default hosted model can be replaced at the prompt if it is unavailable to your Hugging Face account or inference provider.

Project structure

.
|-- documents_project4/  Source documents
|-- ingest.py            Document loading, chunking, embedding, and indexing
|-- rag_chat.py          Retrieval and language-model interaction
|-- main.py              Application entry point
`-- requirements.txt     Python dependencies

Grounding behavior

The assistant is instructed to answer only from retrieved context. When the answer is not present, it returns a fixed fallback response instead of intentionally using outside knowledge. This reduces unsupported answers but does not guarantee factual accuracy; users should verify responses against the displayed sources.

Data attribution

The included demonstration documents summarize public information about Sensata Technologies. Review documents_project4/urls.txt for the source URLs and verify redistribution requirements before publishing the document collection.

Suggested evaluation

Before treating the assistant as production-ready, create a test set containing representative questions, expected source documents, and expected answer facts. Measure retrieval relevance, grounded-answer accuracy, fallback accuracy, latency, and failure behavior.