A Budget-Aware RAG Middleware.

RAGtune transforms static RAG pipelines into dynamic, cost-sensitive feedback loops. It treats cost and latency as first-class constraints, optimizing information gain by iteratively sampling and reranking documents based on real-time feedback.

Explore Docs · View Roadmap · Release Notes

Traditional RAG pipelines are linear (Retrieve → Rerank → Generate) and cost-blind. They either retrieve too little (missing context) or too much (wasting tokens and latency).

RAGtune introduces an Active Learning Loop:

# Clone the repository
git clone https://github.com/yourusername/ragtune.git
cd ragtune

# Install in editable mode
pip install -e .

The fastest way to use RAGtune is via the CLI using declarative YAML or JSON configurations.

Scaffold a new configuration file with default settings.

ragtune init
# Created 'ragtune_config.yaml'

Edit ragtune_config.yaml to define your budget and components.

pipeline:
  budget:
    tokens: 4000
    latency_ms: 1500
  components:
    retriever:
      type: "bm25"
    reranker:
      type: "cross-encoder"

Execute the pipeline instantly from the terminal.

ragtune run ragtune_config.yaml --query "How does Active Learning optimize RAG?"

RAGtune is built on a declarative philosophy. Every pipeline, from simple BM25 to complex iterative estimators, can be described in a single file.

Whether you prefer the readability of YAML or the machine-compatibility of JSON, RAGtune supports both.

YAML Configuration:

# ragtune_config.yaml
pipeline:
  name: "My Pipeline"
  budget: { tokens: 2000 }
  components:
    retriever: { type: "bm25" }

JSON Configuration:

{
  "pipeline": {
    "name": "My Pipeline",
    "budget": { "tokens": 2000 },
    "components": {
      "retriever": { "type": "bm25" }
    }
  }
}

Understand your pipeline flow instantly with ASCII diagrams:

ragtune visualize ragtune_config.yaml

The visualization renders a box-and-arrow diagram representing the pipeline:

╭─────────────────── RAGtune Pipeline: My Pipeline ────────────────────╮
│                                                                      │
│  ┌────────────┐    ┌────────────┐    ┌────────────┐    ┌────────────┐│
│  │ RETRIEVER  │───▶│REFORMULATOR│───▶│  RERANKER  │───▶│ ASSEMBLER  ││
│  ├────────────┤    ├────────────┤    ├────────────┤    ├────────────┤│
│  │ type: bm25 │    │ type: llm  │    │ type: cross│    │ type: greed││
│  └────────────┘    └────────────┘    └────────────┘    └────────────┘│
│                                              ▲                       │
│                    ┌────────────┐    ┌───────┴────┐                  │
│                    │ SCHEDULER  │◀───│  ESTIMATOR │                  │
│                    ├────────────┤    ├────────────┤                  │
│                    │ graceful   │    │ type: base │                  │
│                    └────────────┘    └────────────┘                  │
│                                                                      │
│  Budget: tokens=5000 | rerank_docs=50 | latency_ms=2000              │
╰──────────────────────────────────────────────────────────────────────╯

For a full breakdown of all commands (init, index, validate, run, list, visualize) and the v0.2 configuration schema, see the CLI Reference Guide.

RAGtune sits between your application and your data sources.

graph LR
    User[User Query] --> Controller
    subgraph "RAGtune Middleware"
        Controller -- "Budget Check" --> Tracker[CostTracker]
        Controller -- "Next Batch?" --> Scheduler
        Scheduler -- "Propose" --> Reranker
        Reranker -- "Feedback" --> Estimator
        Estimator -- "Prioritize" --> Scheduler
    end
    Reranker --> FinalContext

See Architecture Docs for deep dives.

RAGtune plays nicely with your existing stack.

• LangChain: Wrap any Retriever as a RAGtune source.
• LlamaIndex: Use QueryEngine or retrievers directly.
• PyTerrier: Full compatibility for IR research benchmarks.

We welcome contributions! Please see our Roadmap to find open tasks.

1. Fork the Project
2. Create your Feature Branch (git checkout -b feature/AmazingFeature)
3. Commit your Changes (git commit -m 'Add some AmazingFeature')
4. Push to the Branch (git push origin feature/AmazingFeature)
5. Open a Pull Request

Distributed under the MIT License. See pyproject.toml for details.

Built with ❤️ by the RAGtune Team.