InferESG: AI-Augmented ESG Analysis

InferESG is an open-source solution that transforms how analysts interact with ESG (Environmental, Social, and Governance) data. Built upon the InferLLM framework, it helps combat greenwashing and validates ESG claims through AI-augmented analysis.

Please Note: Alpha Solution Accelerator Status! This was initially developed as part of a FinTech Scotland Funded research project and whilst it is easy to deploy (see below) and has shown promising results in testing it needs more work (mostly on error handling and presentation) and integration to your data to be most effective. (Contributions to this project and it's parent project InferLLM welcome!)

🎯 Mission

InferESG's mission is to augment ESG analysts' capabilities through intelligent automation while maintaining human expertise and judgment in the loop. Rather than attempting to automate the entire analysis process, InferESG serves as an intelligent assistant that augments human expertise through three key capabilities:

🔍 Comprehensive Report Analysis

• Thorough analysis of sustainability reports
• Step-by-step materiality assessment
• Examination of environmental targets, social initiatives, and governance structures
• Identification of both strong performance areas and potential concerns

💬 Interactive Investigation

• Natural language query interface for ESG data exploration
• Evidence-based responses through multi-agent framework
• Deep dive capabilities into areas of interest or concern
• Cross-referencing claims against external sources

📊 Evidence-Based Analysis

• Clear evidence trails for all conclusions
• Transparent verification paths
• Effective challenge processes
• Support for manual verification

🏗️ Technical Architecture

InferESG employs a sophisticated multi-agent architecture:

• Materiality Agent: Evaluates companies against established frameworks
• Report Agent: Coordinates comprehensive document analysis
• Web Agent: Validates claims against external sources
• Intent Agent: Processes and routes natural language queries
• Validator Agent: Improves response accuracy and completeness

🚀 Key Features

• Integration with key frameworks (TNFD, GRI)
• Materiality assessment capabilities
• Greenwashing detection
• Interactive query system
• Evidence trail generation
• Multi-source validation

📈 Validation & Testing

The system has been validated through rigorous testing:

• Testing across multiple industry sectors
• F1 score metrics for accuracy measurement
• Real-world validation with research industry sponsor
• Cross-validation using multiple AI systems and manual exploratory testing

InferESG Roadmap & Docs

InferESG is based on InferGPT/LLM - more details below:

Want more context about how it works, our roadmap and documentation? Check out the wiki

We are very keen to integration further use of the Knowledge Graph as we believe this will help manage a lot of the risks and improve it's capabilities.

For further reading on InferLLMs's components, see the InferLLM Github

Contribute

We welcome contributions from the community! Whether you're interested in:

• Enhancing the analysis capabilities
• Adding new materiality frameworks
• Improving the agent architecture
• Expanding validation capabilities See the contribution guide for further guidance. Note this guide is in progress!

Getting Started

In the top-right corner of the page, click Fork.

On the next page, select your GitHub account to create the fork under. Wait for the forking process to complete. You now have a copy of the repository in your GitHub account.

Clone the Repository

To clone the repository, you need to have Git installed on your system. Use the official Git installer or follow the terminal commands guide.

Once you have Git installed, follow these steps:

• Open your preferred terminal.
• Navigate to the directory where you want to clone the repository.
• Run the git clone command for the fork you just created.

Initial configuration

There is a example property file .env.example at the root of this project.

Configuration steps:

• Copy the .env.example file at the root of this project.
• Rename the copied file as .env.
• Update the .env file with your wanted configuration following the guidance in the file.

Running the application

There are a few ways to run the application:

1. Docker Compose - run the entire application within Docker containers configured by Docker Compose.
2. Locally - run local instances of the front-end, back-end and a neo4j database.
3. Individual Docker Containers - you may choose to run individual components within a Docker container.

For ease of use, we would recommended that you run the application using Docker Compose.

For instructions on how to run indivdual components locally or within Docker containers, refer to appropriate READMEs:

• frontend README
• backend README
• data README

Running with Docker Compose

• Ensure Docker is installed and running. The easiest way to do this is by using the Docker Desktop app (install it from docker.com if you don't have it).
• In the root directory of the project run docker compose up

  [!WARNING]
  the first time you do this it may take longer as the compose file builds the images.

To re-build the docker images following any changes, run docker compose build first or use docker compose up --build.

For ease of development, after running docker compose build you can run docker compose up --watch. Watch mode allows for the docker container to rebuild when you save a file. This means that it is not required to take down the service after every change to see these implemented. This option is recommended.

• View the frontend at localhost:8650
• View the Neo4j Browser at localhost:7474
• Type the phrase "healthcheck" into the frontend UI to test if everything is connected successfully

Usage

We've designed and tested the current initial workflow with an ESG analyst team. Primarily with the user journey of

• help me assess whether this is a company that we should consider adding to an ESG fund. What claims and potential greenwashing are present?

• What's missing against material issues for the sector?

• Help me (interactively) dig into the details of this company's sustainability claims.

• Save me time wading through the noise to find the right signals to investigate

  Process:

• To get started upload a Sustainability report PDF for an organistion. Be patient as a lot of processing is done initially (plan is to make this more efficient, async and notify the user when done)

• Once the report has been analysed you can click a button to view it in the canvas or download it

• You can then ask questions chat bot style. Behind the scenes InferESG works out whether to use data from the report, web searches, materiality documents, LLM generalist content or a mix

• If you want to see how it's created its output to a question you can click a button and it will (at the time of writing) give you a JSON dump of what agent, queries and source URLs it has used - although the source URLs should also be included in the friendlier UI output

• Suggested questions are provided based on the current conversation context

LICENCE

See LICENCE.md