RAG Pipeline

A command-line Retrieval-Augmented Generation (RAG) pipeline that lets you ingest PDF documents and ask questions against them using LangChain.

Features

• PDF ingestion — loads all PDFs from the data/ directory, splits them into chunks, and stores embeddings in a vector database
• Multi-provider LLM support — choose between Claude, OpenAI, or Gemini as the answering model
• Vector store options — FAISS (local, default) or Pinecone (cloud)
• HuggingFace embeddings — uses all-MiniLM-L6-v2 by default (runs on CPU)

Project Structure

.
├── cli.py           # CLI entrypoint (ingest / ask / chat commands)
├── chain.py         # LLM selection and RetrievalQA chain
├── config.py        # Environment variable loading and defaults
├── embeddings.py    # HuggingFace embedding wrapper
├── loader.py        # PDF loading and text chunking
├── agent.py         # Conversational RAG agent with memory
├── store.py         # Vector store creation, loading, and retrieval
├── init.sh          # Dependency installation script
├── data/            # Place your PDF files here
├── vectorstore/     # FAISS index output (auto-generated)
├── tests/           # Unit tests (pytest)
├── .env.example     # Sample environment configuration
└── .env             # Your local configuration (not committed)

Setup

1. Install dependencies

bash init.sh

2. Configure environment variables

Copy the example env file and fill in your API keys:

cp .env.example .env

Edit .env and set at minimum:

Variable	Description	Default
LLM_PROVIDER	claude, openai, or gemini	claude
LLM_MODEL	Model name for the chosen provider	Provider-specific default
ANTHROPIC_API_KEY	API key (required if provider is claude)	—
OPENAI_API_KEY	API key (required if provider is openai)	—
GOOGLE_API_KEY	API key (required if provider is gemini)	—
VECTOR_STORE	faiss or pinecone	faiss
PINECONE_API_KEY	Required only if using Pinecone	—
PINECONE_INDEX_NAME	Pinecone index name	rag-index
EMBEDDING_MODEL	HuggingFace embedding model	all-MiniLM-L6-v2
CHUNK_SIZE	Characters per text chunk	1000
CHUNK_OVERLAP	Overlap between chunks	200

Usage

Ingest documents

Place your PDF files in the data/ directory, then run:

python cli.py ingest

This loads the PDFs, splits them into chunks, generates embeddings, and saves the vector index.

Ask a question

python cli.py ask "What are the key findings in the report?"

The pipeline retrieves the most relevant chunks and uses the configured LLM to generate an answer, along with source references.

Interactive chat

Start a conversational session with memory across turns:

python cli.py chat

The agent uses the ingested documents as a tool and returns structured JSON responses with confidence levels, source references, and follow-up question suggestions. Type exit or quit to end the session.

Testing

Install pytest (one-time):

pip3 install pytest

Run all tests:

python -m pytest tests/ -v

All external dependencies (LLM APIs, HuggingFace models, FAISS, Pinecone) are mocked, so tests run fast without API keys or model downloads.

Test file	Module	What it covers
test_config.py	config.py	Default values, env var overrides, per-provider model defaults, directory paths
test_loader.py	loader.py	PDF loading and chunking, empty-directory error handling
test_embeddings.py	embeddings.py	Embedding instance creation and configuration
test_store.py	store.py	FAISS/Pinecone create and load paths, retriever with default/custom k
test_chain.py	chain.py	LLM provider selection (Claude/OpenAI/Gemini), unknown provider error, QA chain assembly
test_cli.py	cli.py	Argparse routing, ingest/ask/chat function wiring, output formatting
test_agent.py	agent.py	Pydantic schemas, response parser, RAG tool, agent builder, REPL loop