Modular RAG

A production-oriented, modular Retrieval-Augmented Generation (RAG) framework evolved from an earlier monolithic RAG system with guardrails.

GitHub · Back to Projects

---

• Problem: Monolithic RAG pipeline was hard to extend and debug.
• Approach: Modular components + explicit flow orchestration.
• Outcome: Hybrid retrieval + agentic-ready workflows without refactoring core logic.

---

This project represents a deliberate migration from a tightly coupled, single-pipeline RAG implementation into a modular, composable, and extensible architecture. The goal was to support hybrid retrieval, conditional execution, and future agentic workflows without refactoring core logic.

Key architectural ideas:

• Separation of computation and control — retrieval, embeddings, generation, and reranking are implemented as reusable modules, while execution logic lives in explicit flows.
• Flow-based orchestration — RAG pipelines are defined as executable graphs (sequence, routing, parallelism), enabling simple, hybrid, and agentic-ready workflows.
• Explicit wiring, no magic — predictable startup registration and debuggable flows.
• Hybrid retrieval by design — supports BM25, FAISS dense retrieval, Reciprocal Rank Fusion (RRF), and Maximal Marginal Relevance (MMR).
• Model- and vendor-agnostic — OpenAI-compatible LLM layer supporting local (Ollama) and hosted providers without leaking vendor logic.
• Observability-first — structured logging, tracing, and metrics are built in from the API boundary through flow execution. Each flow step logs: step name, duration, retriever used, top-k doc IDs, and any error category.

Compared to the original monolithic RAG system, this modular approach enables incremental evolution: from simple RAG, to hybrid retrieval, to conditional and agentic workflows — without rewriting existing components.

---

Engineering Notes