README.md

metadata

title: InspectorRAGet
emoji: 🧐
sdk: docker
app_port: 3000
pinned: true

InspectorRAGet

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.

🎥 Demo

🏗️ Build & Deploy

To install and run InspectorRAGet follow the steps below:

Installation

We use yarn as a default package manager.

yarn install

⚠️ node version must be 20.12.0 or higher.

Development server

To start InspectorRAGet in development mode, please run the following command.

yarn dev

Build

To build a static production bundle, please run the following command.

yarn build

Production server

To start InspectorRAGet in production mode, please run the following command.

yarn start

Usage

Once 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.

Use InspectorRAGet through integration notebooks

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

Use InspectorRAGet by manually creating input file

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.

1. Metadata

{
    "name": "Sample experiment name",
    "description": "Sample example description",
    ...

2. Models

    "models": [
      {
        "model_id": "model_1",
        "name": "Model 1",
        "owner": "Model 1 owner",
      },
      {
        "model_id": "model_2",
        "name": "Model 2",
        "owner": "Model 2 owner",
      }
    ],

Notes:

1. Each model must have a unique model_id and name.

3. Metrics

      "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:

1. Each metric must have a unique name.
2. Metric can be of numerical, categorical, or text type.
3. Numerical type metrics must specify range field in [start, end, bin_size] format.
4. Categoricl type metrics must specify values field where each value must have value and numerical_value fields.
5. Text type metric are only accesible in instance level view and not used in any experiment level aggregate statistics and visual elements.

4. Documents

      "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:

1. Each document must have a unique document_id field.
2. Each document must have a text field.

5. Tasks

      "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:

1. Each task must have a unique task_id.
2. Task type can be of rag, or of text_generation, or of chat type.
3. For rag and text_generation type task, input is an array of utterances. An utterance's speaker could be either user or agent. Each utterance must have a text field.
4. For chat type task, input must be array of messages as defined by OpenAI's chat completion APIs (https://platform.openai.com/docs/api-reference/chat/create#chat-create-messages).
5. For rag task, contexts field represents a subset of documents from the documents field relevant to the input and is available to the generative models.
6. targets field is an array of expected gold or reference texts.
7. category is an optional field that represents the type of task for grouping similar tasks.
8. filters is a top-level field (parallel to tasks) which specifies an array of fields defined inside tasks for filtering tasks during analysis.

6. Evaluations

"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:

1. evaluations field must contain evaluation for every model defined in models section and on every task in tasks section. Thus, total number of evaluations is equal to number of models (M) X number of tasks (T) = M X T
2. Each evaluation must be associated with single task and single model.
3. Each evaluation must have model prediction on a task captured in the model_response field.
4. annotations field captures ratings on the model for a given task and for every metric specified in the metrics field.
5. Each metric annotation is a dictionary containing worker ids as keys. In the example above, system is a worker id.
6. Annotation from any worker on all metrics must be in the form of a dictionary. At minimum, such dictionary contains value key capturing model's rating for the metric by the worker.

Citation

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}
}