RAGLight

A lightweight, typed Python framework for classic and agentic retrieval-augmented generation.

Build a local RAG in minutes

RAGLight keeps retrieval-augmented generation explicit: choose an LLM, an embedding model, and a vector store, then index a folder and ask a question.

Quickstart

With Ollama running locally, install RAGLight and pull the two models:

uv add "raglight[qdrant,ollama,documents]"
ollama pull nomic-embed-text:latest
ollama pull qwen3:0.6b

Put your .txt, .md, .html, source-code, or PDF files in ./documents, then save this as local_rag.py:

from raglight import RAGPipeline
from raglight.integrations import default_registries
from raglight.integrations.ollama import OllamaEmbeddings, OllamaLLM
from raglight.integrations.qdrant import QdrantVectorStore
from raglight.models import FolderSource

embeddings = OllamaEmbeddings(model_name="nomic-embed-text:latest")
llm = OllamaLLM(model_name="qwen3:0.6b", preload_model=False)
vector_store = QdrantVectorStore(
    collection_name="documents",
    embeddings_model=embeddings,
    persist_directory="./raglight-db",
)

rag = RAGPipeline(
    llm=llm,
    embeddings=embeddings,
    vector_store=vector_store,
    document_processors=default_registries.document_processors,
)

ingestion = rag.ingest([FolderSource("./documents")])
if ingestion.failures:
    raise RuntimeError(ingestion.to_dict())

response = rag.generate("What are the main topics in these documents?")

print(response.answer)
print("Sources:", [source.metadata for source in response.sources])

Run it:

uv run python local_rag.py

Qdrant runs locally in ./raglight-db; no Qdrant server or API key is needed. The ollama extra includes both the LLM and embedding adapters. The documents extra provides the built-in file processors.

Explore the framework

Classic RAG

Retrieve, optionally rerank, and generate a structured answer.

Agentic RAG

Run a bounded tool loop through a separate public API.

Directory ingestion

Recursively process folders with extension-based document processors.

Bring your own components

Implement structural protocols without inheriting from framework classes.

Get started

Install local models, ingest a directory, and ask your first question.

Copy-paste examples

Browse runnable Ollama, OpenAI, custom endpoint, Agentic RAG, MCP, and extension examples.