InspectorRAGet, an introspection platform for RAG evaluation. InspectorRAGet allows the user to analyze aggregate and instance-level performance of RAG systems, using both human and algorithmic metrics as well as annotator quality.
InspectorRAGet has been developed as a React web application built with NextJS 14 framework and the Carbon Design System.
To install and run InspectorRAGet follow the steps below:
We use yarn as a default package manager.
yarn install20.12.0 or higher.
To start InspectorRAGet in development mode, please run the following command.
yarn devTo build a static production bundle, please run the following command.
yarn devTo start InspectorRAGet in production mode, please run the following command.
yarn startOnce you have started InspectorRAGet, the next step is import a json file with the evaluation results in the format expected by the platform. You can do this in two ways:
- Use one of our integration notebooks, showing how to use InspectorRAGet in combination with popular evaluation frameworks.
- Manually convert the evaluation results into the expected format by consulting the documentation of InspectorRAGet's file format.
To make it easier to get started, we have created notebooks showcasing how InspectorRAGet can be used in combination with popular evaluation frameworks. Each notebook demonstrates how to use the corresponding framework to run an evaluation experiment and transform its output to the input format expected by InspectorRAGet for analysis. We provide notebooks demonstrating integrations of InspectorRAGet with the following popular frameworks:
| Framework | Description | Integration Notebook |
|---|---|---|
| Language Model Evaluation Harness | Popular evaluation framework used to evaluate language models on different tasks | LM_Eval_Demonstration.ipynb |
| Ragas | Popular evaluation framework specifically designed for the evaluation of RAG systems through LLM-as-a-judge techniques | Ragas_Demonstration.ipynb |
| HuggingFace | Offers libraries and assets (incl. datasets, models, and metric evaluators) that can be used to both create and evaluate RAG systems | HuggingFace_Demonstration.ipynb |
If you want to use your own code/framework, not covered by the integration notebooks above, to run the evaluation, you can manually transform the evaluation results to the input format expected by InspectorRAGet, described below. Examples of input files in the expected format can be found in the data folder.
The experiment results json file expected by InspectorRAGet can be broadly split into six sections along their functional boundaries. The first section captures general details about the experiment in name, description and timestamp fields. The second and third sections describe the
sets of models and metrics used in the experiment via the models and metrics fields, respectively. The last three sections cover the dataset and the outcome of evaluation experiment in the form of documents, tasks and evaluations fields.
{
"name": "Sample experiment name",
"description": "Sample example description",
... "models": [
{
"model_id": "model_1",
"name": "Model 1",
"owner": "Model 1 owner",
},
{
"model_id": "model_2",
"name": "Model 2",
"owner": "Model 2 owner",
}
],Notes:
- Each model must have a unique
model_idandname.
"numerical": [
{
"name": "metric_a",
"display_name": "Metric A",
"description": "Metric A description",
"author": "algorithm | human",
"type": "numerical",
"aggregator": "average",
"range": [0, 1, 0.1]
},
{
"name": "metric_b",
"display_name": "Metric B",
"description": "Metric B description",
"author": "algorithm | human",
"type": "categorical",
"aggregator": "majority | average",
"values": [
{
"value": "value_a",
"display_value": "A",
"numeric_value": 1
},
{
"value": "value_b",
"display_value": "B",
"numeric_value": 0
}
]
},
{
"name": "metric_c",
"display_name": "Metric C",
"description": "Metric C description",
"author": "algorithm | human",
"type": "text"
}
],Notes:
- Each metric must have a unique name.
- Metric can be of
numerical,categorical, ortexttype. - Numerical type metrics must specify
rangefield in[start, end, bin_size]format. - Categoricl type metrics must specify
valuesfield where each value must havevalueandnumerical_valuefields. - Text type metric are only accesible in instance level view and not used in any experiment level aggregate statistics and visual elements.
"documents": [
{
"document_id": "GUID 1",
"text": "document text 1",
"title": "document title 1"
},
{
"document_id": "GUID 2",
"text": "document text 2",
"title": "document title 2"
},
{
"document_id": "GUID 3",
"text": "document text 3",
"title": "document title 3"
}
],Notes:
- Each document must have a unique
document_idfield. - Each document must have a
textfield.
"filters": ["category"],
"tasks": [
{
"task_id": "task_1",
"task_type": "rag",
"category": "grounded",
"input": [
{
"speaker": "user",
"text": "Sample user query"
}
],
"contexts": [
{
"document_id": "GUID 1"
}
],
"targets": [
{
"text": "Sample response"
}
]
},
{
"task_id": "task_2",
"task_type": "rag",
"category": "random",
"input": [
{
"speaker": "user",
"text": "Hello"
}
],
"contexts": [
{
"document_id": "GUID 2"
}
],
"targets": [
{
"text": "How can I help you?"
}
]
}
],Notes:
- Each task must have a unique
task_id. - Task type can be of
question_answering,conversation, or ofragtype. inputis an array of utterances. An utterance's speaker could be eitheruseroragent. Each utterance must have atextfield.contextsfield represents a subset of documents from thedocumentsfield relevant to theinputand is available to the generative models.targetsfield is an array of expected gold or reference texts.categoryis an optional field that represents the type of task for grouping similar tasks.filtersis a top-level field (parallel totasks) which specifies an array of fields defined insidetasksfor filtering tasks during analysis.
"evaluations": [
{
"task_id": "task_1 | task_2",
"model_id": "model_1 | model_2",
"model_response": "Model response",
"annotations": {
"metric_a": {
"system": {
"value": 0.233766233766233
}
},
"metric_b": {
"system": {
"value": "value_a | value_b"
}
},
"metric_c": {
"system": {
"value": "text"
}
},
}
}
]Notes:
evaluationsfield must contain evaluation for every model defined inmodelssection and on every task intaskssection. Thus, total number of evaluations is equal to number of models (M) X number of tasks (T) = M X T- Each evaluation must be associated with single task and single model.
- Each evaluation must have model prediction on a task captured in the
model_responsefield. annotationsfield captures ratings on the model for a given task and for every metric specified in themetricsfield.- Each metric annotation is a dictionary containing worker ids as keys. In the example above,
systemis a worker id. - Annotation from any worker on all metrics must be in the form of a dictionary. At minimum, such dictionary contains
valuekey capturing model's rating for the metric by the worker.
If you use InspectorRAGet in your research, please cite our paper:
@misc{fadnis2024inspectorraget,
title={InspectorRAGet: An Introspection Platform for RAG Evaluation},
author={Kshitij Fadnis and Siva Sankalp Patel and Odellia Boni and Yannis Katsis and Sara Rosenthal and Benjamin Sznajder and Marina Danilevsky},
year={2024},
eprint={2404.17347},
archivePrefix={arXiv},
primaryClass={cs.SE}
}