RAG with LangChain

A production-oriented Retrieval-Augmented Generation (RAG) system built using LangChain and Clean Architecture.

Unlike my previous RAG implementation—which was built entirely from scratch without any framework—this project explores how to build a maintainable and extensible RAG using LangChain while keeping framework-specific code isolated from the core domain.

Goals

• Clean Architecture
• Framework-agnostic domain layer
• Modular ingestion pipeline
• Local-first stack
• Production-oriented design
• Easy experimentation with retrieval techniques

---

Current Architecture

                PDF
                 │
          PDF Loader
                 │
        Document Processors
                 │
        Recursive Chunker
                 │
          Ollama Embeddings
                 │
      PostgreSQL + pgvector
                 │
──────────── Retrieval ────────────

Question
     │
PGVector Retriever
     │
Cross-Encoder Reranker
     │
Relevant Context
     │
Prompt Builder
     │
Llama 3.1 (Ollama)
     │
Answer

---

Project Structure

rag/
│
├── app/
│   ├── api/
│   ├── application/
│   ├── config/
│   ├── domain/
│   └── infrastructure/
│
├── evaluation/
├── tests/
├── main.py
└── pyproject.toml

---

Tech Stack

LLM

• Ollama
• Llama 3.1 8B

Embeddings

• nomic-embed-text

Vector Database

• PostgreSQL
• pgvector

Framework

• LangChain

Retrieval

• Dense Retrieval
• Cross-Encoder Reranking

Evaluation

• DeepEval
• RAGAS

---

Current Features

• PDF ingestion
• Recursive text chunking
• Custom document processing pipeline
• Ollama embeddings
• PostgreSQL + pgvector storage
• Dense semantic retrieval
• Cross-encoder reranking
• Modular retrieval and generation services
• Clean Architecture

---

Planned Features

• Hybrid retrieval (Dense + BM25)
• Metadata filtering
• Context compression
• Query rewriting
• Multi-query retrieval
• Reciprocal Rank Fusion (RRF)
• LangGraph workflows
• Streaming responses
• Source citations
• FastAPI API
• LangSmith tracing
• Comprehensive evaluation benchmarks

---

Design Principles

• Keep business logic independent of LangChain.
• Depend on interfaces instead of implementations.
• Isolate infrastructure components.
• Make every component replaceable.
• Prefer composition over inheritance.

---

Related Project

This repository is the LangChain counterpart to my previous project, where every RAG component—including embeddings, vector search, prompt orchestration, and retrieval—was implemented manually without using LangChain.