Quivr — Opinionated RAG Framework for Any LLM

What it is

Quivr is a Python RAG (retrieval-augmented generation) framework that takes an opinionated approach to building knowledge bases from documents. You feed it files (PDFs, text, markdown, and more), and Quivr handles ingestion, chunking, embedding, and retrieval. Then you query your document collection in natural language using any LLM backend. The framework is designed to get a working RAG pipeline running in minutes, not days.

Developers and teams who need to build document Q&A systems, internal knowledge bases, or AI assistants grounded in specific documents benefit from Quivr. Its opinionated defaults mean less configuration compared to more modular frameworks.

How it saves time or tokens

Quivr's opinionated design eliminates the decision fatigue of choosing chunking strategies, embedding models, and retrieval methods. The defaults work well for most document types. By handling the entire RAG pipeline in a few lines of code, Quivr saves the days of setup that more flexible frameworks require. The token_estimate for this workflow is approximately 337 tokens for a basic query.

How to use

1. Install Quivr via pip
2. Create a Brain from your document files
3. Ask questions in natural language and get grounded answers

Example

from quivr_core import Brain

# Create a brain from your documents
brain = Brain.from_files(
    name='my-knowledge-base',
    file_paths=['./report.pdf', './notes.md']
)

# Ask questions
answer = brain.ask('What were the key findings?')
print(answer.answer)
print(answer.sources)  # Shows which documents were used

Related on TokRepo

• RAG tools — Compare RAG frameworks and retrieval solutions
• AI tools for documents — Browse document processing and parsing tools

Common pitfalls

• Large PDF files with complex layouts may chunk poorly; preprocess scanned documents with OCR before ingestion
• The default embedding model requires an API key; configure a local embedding model for fully offline operation
• Quivr's opinionated defaults work well for general documents but may need tuning for highly specialized technical content