RAG Pipeline Builder

A production-ready Retrieval-Augmented Generation pipeline that lets you chat with your own documents. Ingest PDFs, Markdown, text files, or URLs — then query them with a grounded LLM that cites its sources.

Architecture

Documents (PDF / TXT / MD / URL)
        │
        ▼
  Load & chunk          RecursiveCharacterTextSplitter
        │                 chunk_size=800, overlap=120
        ▼
  Embed & store         text-embedding-3-small → Chroma (local)
        │                                      → Pinecone (cloud)
        ▼
  Retrieve (MMR)        Top-K diverse chunks via Max Marginal Relevance
        │
        ▼
  Rerank                cross-encoder/ms-marco-MiniLM-L-6-v2
        │                 narrows to top-N most relevant chunks
        ▼
  Generate              GPT-4o / Claude — grounded answer + source citations

Requirements

Python >= 3.10
langchain
langchain-openai
langchain-community
chromadb
sentence-transformers
fastapi
uvicorn
streamlit
pypdf
tiktoken
python-dotenv
requests

Install everything:

pip install langchain langchain-openai langchain-community \
            chromadb sentence-transformers fastapi uvicorn \
            streamlit pypdf tiktoken python-dotenv requests

Setup

# 1. Clone or copy the files
cp .env.example .env
2. Add your OpenAI key to .env
OPENAI_API_KEY=sk-...
3. Ingest your documents
python rag_pipeline.py ingest --source ./docs
4. Ask a question
python rag_pipeline.py query --question "What is our refund policy?"

Usage

CLI

# Ingest a directory of PDFs and Markdown files
python rag_pipeline.py ingest --source ./docs
Ingest a single file
python rag_pipeline.py ingest --source ./contracts/tos.pdf
Ingest a web page
python rag_pipeline.py ingest --source https://docs.yourproduct.com/faq
Query (with reranker, default)
python rag_pipeline.py query --question "What are the SLA guarantees?"
Query without reranker (faster, slightly less precise)
python rag_pipeline.py query --question "Who do I contact for billing?" --no-reranker
Stream the answer token by token
python rag_pipeline.py query --question "Summarise the privacy policy" --stream

Streamlit chat UI

streamlit run rag_pipeline.py

Opens a browser chat interface with a sidebar for ingesting new documents on the fly.

REST API

uvicorn rag_pipeline:app --reload

Example request:

curl -X POST http://localhost:8000/query \
  -H "Content-Type: application/json" \
  -d '{"question": "What is our cancellation policy?", "use_reranker": true}'

Example response:

{
  "question": "What is our cancellation policy?",
  "answer": "Customers may cancel within 30 days for a full refund...\n\nSources:\n[1] tos.pdf, p.4\n[2] faq.md"
}

Configuration

All settings live in .env. Key options:

Swapping to Pinecone

# In rag_pipeline.py, replace build_vectorstore() with:
from langchain_community.vectorstores import Pinecone
import pinecone
pinecone.init(api_key=os.getenv("PINECONE_API_KEY"), environment=os.getenv("PINECONE_ENV"))
vs = Pinecone.from_documents(chunks, embeddings, index_name=os.getenv("PINECONE_INDEX"))

Using local models (Ollama)

ollama pull nomic-embed-text
ollama pull mistral

# Replace embeddings + LLM in rag_pipeline.py:
from langchain_community.embeddings import OllamaEmbeddings
from langchain_community.chat_models import ChatOllama
embeddings = OllamaEmbeddings(model="nomic-embed-text")
llm        = ChatOllama(model="mistral", temperature=0)

Fully local, zero API costs.

Output example

> Q: What is the enterprise refund window?
> A: Enterprise customers are entitled to a full refund within 30 days of
purchase, provided the cancellation request is submitted via the billing
portal. Partial refunds are not available after the 30-day window.
Sources:
[1] enterprise-tos.pdf, p.4
[2] pricing-faq.md

Project structure

rag_pipeline.py      Main pipeline — ingest, retrieve, generate, API, UI
.env.example         Environment variable template
.env                 Your local config (git-ignored)
chroma_db/           Persisted vector store (auto-created on first ingest)
docs/                Put your source documents here

Delivering this as a freelance gig

A typical engagement looks like this:

1. Discovery call — understand the client's document types, volume, and query patterns
2. Ingest + test — run their docs through the pipeline, tune CHUNK_SIZE and TOP_K
3. Reranker eval — compare answers with/without the cross-encoder on 20 sample questions
4. Handoff — deliver the code, .env.example, and a short Loom walkthrough

Upsell hooks: persistent chat history (add Redis), multi-user auth (add Clerk/Auth0), Slack bot integration, automated re-ingestion on doc update (add a webhook).