I built ragway — a Python RAG library controlled by a single YAML file

Swapanth Vakapalli — Mon, 16 Mar 2026 11:07:33 +0000

I spent the last few months building ragway, a Python RAG library where the entire pipeline is controlled by a single YAML file.

The Problem

Every RAG project I worked on ended up with the same messy code: branching logic to switch between LLMs, copy-pasted vectorstore setup, tightly coupled retrieval and generation. When requirements changed — swap GPT-4 for Llama, move from FAISS to Pinecone — it meant touching a lot of code that shouldn't need to change.

The Solution

ragway separates configuration from code completely. One YAML file describes your entire pipeline:

version: "1.0"
pipeline: hybrid
plugins:
  llm:
    provider: groq
    model: llama-3.3-70b-versatile
    api_key: ${GROQ_API_KEY}
  embedding:
    provider: bge
    model: BAAI/bge-large-en-v1.5
  vectorstore:
    provider: qdrant
    index_path: .ragway/index
  reranker:
    enabled: true
    provider: bge
  chunking:
    strategy: recursive
    chunk_size: 512

Your Python code never changes:

from ragway import RAG
import asyncio

rag = RAG.from_config("rag.yaml")

async def main():
    count = await rag.ingest("./docs/")
    print(f"Ingested {count} chunks")

    answer = await rag.query("What is in my documents?")
    print(answer)

asyncio.run(main())

Switch the LLM without re-ingesting:

fast = rag.switch(llm="groq")

What's Configurable

Component	Options
LLM	`anthropic`, `openai`, `mistral`, `groq`, `llama`, `local`
Vectorstore	`faiss`, `chroma`, `qdrant`, `pinecone`, `weaviate`, `pgvector`
Pipeline	`naive`, `hybrid`, `self`, `long_context`, `agentic`
Reranker	`cohere`, `bge`, `cross_encoder` (or none)
Chunking	`fixed`, `recursive`, `semantic`, `sliding_window`, `hierarchical`
Embedding	`openai`, `bge`, `cohere`, `sentence_transformer`

Running Fully Local (Zero API Costs)

pip install ragway[faiss,bge]

plugins:
  llm:
    provider: llama
    model_path: ./models/llama-3.1-8b.gguf
  embedding:
    provider: bge
    model: BAAI/bge-large-en-v1.5
  vectorstore:
    provider: faiss
    index_path: .ragway/index
  reranker:
    enabled: false

Under the Hood

Python 3.11, pydantic v2, fully async I/O
ComponentRegistry resolves YAML strings to real classes
ConfigLoader handles ${ENV_VAR} substitution and key injection
357 tests, mypy strict mode, ruff clean
Optional dependencies — install only what you use

Interactive Config Builder

The docs site has a visual config builder where you pick your providers and download a ready-to-use rag.yaml instantly.

ragway.vercel.app

DEV Community: Swapanth Vakapalli