open-rag-eval

Open RAG Eval

Evaluate and improve your Retrieval-Augmented Generation (RAG) pipelines with open-rag-eval, an open-source Python evaluation toolkit.

Evaluating RAG quality can be complex. open-rag-eval provides a flexible and extensible framework to measure the performance of your RAG system, helping you identify areas for improvement. Its modular design allows easy integration of custom metrics and connectors for various RAG implementations.

Importantly, open-rag-eval's metrics do not require golden chunks or golden answer, making RAG evaluation easy and scalable. This is achieved by utilizing UMBRELA and AutoNuggetizer, techniques originating and researched in Jimmy Lin's lab at UWaterloo.

Out-of-the-box, the toolkit includes:

• An implementation of the evaluation metrics used in the TREC-RAG benchmark.
• A connector for the Vectara RAG platform.
• Connectors for LlamaIndex and LangChain (more coming soon...)

Key Features

• Standard Metrics: Provides TREC-RAG evaluation metrics ready to use.
• Modular Architecture: Easily add custom evaluation metrics or integrate with any RAG pipeline.
• Detailed Reporting: Generates per-query scores and intermediate outputs for debugging and analysis.
• Visualization: Compare results across different configurations or runs with plotting utilities.

Getting Started Guide

This guide walks you through an end-to-end evaluation using the toolkit. We'll use Vectara as the example RAG platform and the TRECRAG evaluator.

Prerequisites

• Python: Version 3.9 or higher.
• OpenAI API Key: Required for the default LLM judge model used in some metrics. Set this as an environment variable: export OPENAI_API_KEY='your-api-key'
• Vectara Account: To enable the Vectara connector, you need:
  • A Vectara account.
  • A corpus containing your indexed data.
  • An API key with querying permissions.
  • Your Customer ID and Corpus key.

Installation

In order to build the library from source, which is the recommended method to follow the sample instructions below you can do:

$ git clone https://github.com/vectara/open-rag-eval.git
$ cd open-rag-eval
$ pip install -e .

If you want to install directly from pip, which is the common method if you want to use the library in your own pipeline instead of running the samples, you can run:

pip install open-rag-eval

After installing the library you can follow instructions below to run a sample evaluation and test out the library end to end.

Using Open RAG Eval with the Vectara connector

Step 1. Define Queries for Evaluation

Create a CSV file that contains the queries (for example queries.csv), which contains a single column named query, with each row representing a query you want to test against your RAG system.

Example queries file:

query
What is a blackhole?
How big is the sun?
How many moons does jupiter have?

Step 2. Configure Evaluation Settings

Edit the eval_config_vectara.yaml file. This file controls the evaluation process, including connector options, evaluator choices, and metric settings.

• Ensure your queries file is listed under input_queries, and fill in the correct values for generated_answers and eval_results_file
• Choose an output folder (where all artifacts will be stored) and put it unde results_folder
• Update the connector section (under options/query_config) with your Vectara corpus_key.
• Customize any Vectara query parameter to tailor this evaluation to a query configuration set.

In addition, make sure you have VECTARA_API_KEY and OPENAI_API_KEY available in your environment. For example:

• export VECTARA_API_KEY='your-vectara-api-key'
• export OPENAI_API_KEY='your-openai-api-key'

Step 3. Run evaluation!

With everything configured, now is the time to run evaluation! Run the following command from the root folder of open-rag-eval:

python open_rag_eval/run_eval.py --config config_examples/eval_config_vectara.yaml

You should see the evaluation progress on your command line. Once it's done, detailed results will be saved to a local CSV file (in the file listed under eval_results_file) where you can see the score assigned to each sample along with intermediate output useful for debugging and explainability.

Note that a local plot for each evaluation is also stored in the output folder, under the filename listed as metrics_file.

Step 4. Visualize results

You can use the plot_results.py script to plot results from your eval runs. Multiple different runs can be plotted on the same plot allowing for easy comparison of different configurations or RAG providers:

To plot one result:

python open_rag_eval/plot_results.py results.csv

Or to plot multiple results:

python open_rag_eval/plot_results.py results_1.csv results_2.csv results_3.csv

By default the run_eval.py script will plot metrics and save them to the results folder.

Using Open RAG Eval with your own RAG outputs

If you are using RAG outputs from your own pipeline, make sure to put your RAG output in a format that is readable by the toolkit (See data/test_csv_connector.csv as an example).

Step 1. Configure Evaluation Settings

Copy vectara_eval_config.yaml to xxx_eval_config.yaml (where xxx is the name of your RAG pipeline) as follows:

• Comment out or delete the connector section
• Ensure input_queries, results_folder, generated_answers and eval_results_file are properly configured. Specifically the generated answers need to exist in the results folder.

Step 2. Run evaluation!

With everything configured, now is the time to run evaluation! Run the following command:

python open_rag_eval/run_eval.py --config xxx_eval_config.yaml

and you should see the evaluation progress on your command line. Once it's done, detailed results will be saved to a local CSV file where you can see the score assigned to each sample along with intermediate output useful for debugging and explainability.

Step 3. Visualize results

You can use the open_rag_eval/plot_results.py script to plot results from your eval runs. Multiple different runs can be plotted on the same plot allowing for easy comparison of different configurations or RAG providers. For example if the output evaluation results from two runs are saved in open_eval_results_1.csv and open_eval_results_2.csv you can plot both of them as follows:

python open_rag_eval/plot_results.py results_1.csv results_2.csv

Visualization: Deep dive into results

The visualization in the steps abopve shows you the aggregated metrics across one or more runs of the evaluation on several queries. If you want to deep dive into the results, we have a results viewer which enables easy viewing od the produced metrics CSV where you can look at the intermediate results and detailed breakdown of scores and metrics on a per query basis. To do this:

cd open_rag_eval/viz/
streamlit run visualize.py

Note that you will need to have streamlit installed in your environment (which should be the case if you've installed open-rag-eval). Once you upload your evaluation results CSV (results.csv by default) you can select a query to view detailed metrics for such as the produced nuggets by the AutoNuggetizer, the UMBRELA scores assigned to each retrieved result and so on.

How does open-rag-eval work?

Evaluation Workflow

The open-rag-eval framework follows these general steps during an evaluation:

• (Optional) Data Retrieval: If configured with a connector (like the Vectara connector), call the specified RAG provider with a set of input queries to generate answers and retrieve relevant document passages/contexts. If using pre-existing results (input_results), load them from the specified file.
• Evaluation: Use a configured Evaluator to assess the quality of the RAG results (query, answer, contexts). The Evaluator applies one or more Metrics.
• Scoring: Metrics calculate scores based on different quality dimensions (e.g., faithfulness, relevance, context utilization). Some metrics may employ judge Models (like LLMs) for their assessment.
• Reporting: Generate a detailed report (typically CSV) containing the scores for each query, along with intermediate data useful for analysis and debugging.

Core Abstractions

• Metrics: Metrics are the core of the evaluation. They are used to measure the quality of the RAG system, each metric has a different focus and is used to evaluate different aspects of the RAG system. Metrics can be used to evaluate the quality of the retrieval, the quality of the (augmented) generation, the quality of the RAG system as a whole.
• Models: Models are the underlying judgement models used by some of the metrics. They are used to judge the quality of the RAG system. Models can be diverse: they may be LLMs, classifiers, rule based systems, etc.
• Evaluators: Evaluators can chain together a series of metrics to evaluate the quality of the RAG system.
• RAGResults: Data class representing the output of a RAG pipeline for a single query (input query, generated answer, retrieved contexts/documents). This is the primary input for evaluation.
• ScoredRAGResult: Data class holding the original RAGResults plus the scores assigned by the Evaluator and its Metrics. These are typically collected and saved to the output report file.

Web API

For programmatic integration, the framework provides a Flask-based web server.

Endpoints:

• /api/v1/evaluate: Evaluate a single RAG output provided in the request body.
• /api/v1/evaluate_batch: Evaluate multiple RAG outputs in a single request.

Run the Server:

python open_rag_eval/run_server.py

See the API README for detailed documentation for the API.

About Connectors

Open-RAG-Eval uses a plug-in connector architecture to enable testing various RAG platforms. Out of the box it includes connectors for Vectara, LlamaIndex and Langchain.

Here's how connectors work:

• All connectors are derived from the Connector class, and need to define the fetch_data method.
• The Connector class has a utility method called read_queries which is helpful in reading the input queries.
• When implementing fetch_data you simply go through all the queries, one by one, and call the RAG system with that query.
• The output is stored in the results file, with a N rows per query, where N is the number of passages (or chunks) including these fields
  • query_id: a unique ID for the query
  • query text: the actual query text string
  • passage: the passage (aka chunk)
  • passage_id: a unique ID for this passage (you can use just the passage number as a string)
  • generated_answer: text of the generated response or answer from your RAG pipeline, including citations in [N] format.

See the example results file for an example results file

All 3 existing connectors (Vectara, Langchain and LlamaIndex) provide a good reference for how to implement a connector.

Author

👤 Vectara

• Website: vectara.com
• Twitter: @vectara
• GitHub: @vectara
• LinkedIn: @vectara
• Discord: @vectara

🤝 Contributing

Contributions, issues and feature requests are welcome and appreciated!

Feel free to check issues page. You can also take a look at the contributing guide.

Show your support

Give a ⭐️ if this project helped you!

📝 License

Copyright © 2025 Vectara.
This project is Apache 2.0 licensed.