LangChain RAG Demo (FAISS + Local Embeddings)

A minimal Retrieval-Augmented Generation (RAG) demo using LangChain, FAISS (local vector store), and HuggingFace embeddings.
Optional UI via Streamlit. Works with OpenAI or Ollama (local LLM) as the generator.

Designed as a fast portfolio project to show orchestration skills (LangChain / LlamaIndex-equivalent) for production-oriented roles.

Features

• Document ingestion: chunking + local embeddings (sentence-transformers/all-MiniLM-L6-v2).
• Vector store: FAISS saved locally under ./storage/faiss_index.
• RAG pipeline: retrieve top-k chunks and answer with cited sources.
• Two modes:
  • CLI: rag_query.py --question "...".
  • Streamlit app: streamlit run src/app.py.
• Pluggable LLM provider: OpenAI or Ollama.

Quickstart

1) Create environment

python -m venv .venv
source .venv/bin/activate  # Windows: .venv\Scripts\activate
pip install -r requirements.txt
cp .env.example .env

2) Configure LLM (pick one)

• OpenAI: set LLM_PROVIDER=openai and OPENAI_API_KEY=... in .env.
• Ollama (local): set LLM_PROVIDER=ollama and (optionally) OLLAMA_MODEL=llama3.1 (or similar). Install Ollama and pull a model, e.g. ollama pull llama3.1.

3) Ingest documents

python src/ingest.py --docs_dir data/sample_docs --persist_dir storage/faiss_index

4) Query (CLI)

python src/rag_query.py --persist_dir storage/faiss_index --question "What is RAG and why is chunking useful?"

5) Streamlit UI (optional)

streamlit run src/app.py

Project Structure

langchain-rag-demo/
├─ README.md
├─ requirements.txt
├─ .gitignore
├─ .env.example
├─ data/
│  └─ sample_docs/
│     ├─ 1_intro.txt
│     ├─ 2_rag.txt
│     └─ 3_eval.txt
├─ storage/            # created after ingestion
└─ src/
   ├─ ingest.py
   ├─ rag_query.py
   ├─ app.py
   └─ utils.py

Notes

• Uses local embeddings to avoid external dependency for retrieval.
• LLM can be swapped; defaults to OpenAI if LLM_PROVIDER is unset and an API key exists, otherwise tries Ollama.
• For a cloud vector DB (Pinecone/Qdrant), swap FAISS with the appropriate LangChain vector store in ingest.py/rag_query.py.

License

MIT

---

LlamaIndex Variant

We also include a llamaindex_demo/ directory with a minimal LlamaIndex-based RAG pipeline.

Run LlamaIndex Demo

python llamaindex_demo/basic_rag.py --question "What is RAG?"

---

NEW: LlamaIndex + Pinecone/Qdrant Variants

This project now includes:

• LlamaIndex demo (local FAISS) with CLI and Streamlit.
• Vector DB backends for LangChain: Pinecone and Qdrant.

LlamaIndex (local FAISS)

pip install -r requirements.txt
python llamaindex_demo/src/ingest.py --docs_dir data/sample_docs --persist_dir storage/faiss_index_llama
python llamaindex_demo/src/query.py --persist_dir storage/faiss_index_llama --question "What is RAG?"
# UI:
streamlit run llamaindex_demo/src/app.py

LangChain + Pinecone

# Env (.env):
# PINECONE_API_KEY=...
# PINECONE_INDEX=rag-demo-index
python src/ingest_pinecone.py --docs_dir data/sample_docs --index_name $PINECONE_INDEX
python src/rag_query_pinecone.py --index_name $PINECONE_INDEX --question "Explain chunking"

LangChain + Qdrant

# Env (.env):
# QDRANT_URL=http://localhost:6333
# QDRANT_API_KEY=  # if cloud/secured
# QDRANT_COLLECTION=rag_demo
python src/ingest_qdrant.py --docs_dir data/sample_docs --collection $QDRANT_COLLECTION
python src/rag_query_qdrant.py --collection $QDRANT_COLLECTION --question "How to evaluate RAG?"

Why this matters for production

• Reproducibility: scripted ingestion, deterministic chunking, and versioned embeddings.
• Observability: JSONL metrics logging (latency, token estimate, source count) + reports/eval_report.md.
• Swap-in backends: FAISS local for dev; Pinecone/Qdrant for scale without code rewrites.
• Portability: Dockerfile + Makefile targets; easy to run locally or in CI.