This is the twentieth conversation of the 100+ Conversations to Inspire Our New Direction (#OKFN100) project.
To guide our coming years, we are meeting with over 100 people to discuss the future of open knowledge, shaped by a diverse set of visions from artists, activists, academics, archivists, thinkers, policymakers, data scientists, educators, and community leaders from around the world.
How can openness accelerate and strengthen struggles against the complex challenges of our time? This is the key question behind conversations like the one you can read below.
In today’s conversation, we speak to Biotechnologist Renu Kumari and Ph.D. Scholar Udita Agarwal, members of the citizen science initiative #semanticClimate, and leaders of the development team of the soon-to-be-launched Climate Academy Chatbot (CABot) .
CABot is the first prototype to emerge from the OKFN’s AI Learning Labs project. In each Lab, people and organisations collaborate to learn together and propose solutions to research questions relating to AI. These results are then shared openly for reuse by others.
In this case, the Climate Academy in Europe was looking to expand its educational offering to young people and schools – with a chatbot. But it had to be in a responsible and climate-friendly manner, grounded in pre-approved literature, datasets and conversational guardrails.
They teamed up with #semanticClimate to build a prototype.
In this interview, Renu Kumari and Udita Agarwal – both based in Delhi, India – share the key design decisions behind CABot. The entire codebase is available and documented on GitHub.
How did they do it?
You can find out in the conversation below, which took place on 25 June 2026.
Lucas Pretti: The language surrounding AI today is contentious. Terms such as ‘ethical’, ‘public interest’, ‘open’ and ‘trustworthy’, are so often used for marketing of corporate tech and ‘openwashing’ they are rendered somewhat meaningless. You describe The Climate Academy Chatbot as a ‘responsible’ chatbot. What does that mean in practice?
Udita Agarwal: Before building the system, we spent a lot of time asking ourselves what ‘responsible’ actually means. For us, the most important aspect is trustworthiness.
When a user receives information from the chatbot, they should be able to trust that information and trace it back to a source. The chatbot shouldn’t generate content simply because it has been asked a question. It should only provide answers that come from a reliable source.
To achieve that, we created our own knowledge base. The chatbot’s answers come only from sources we have selected and curated ourselves: The Climate Academy book, and a climate encyclopedia that we built at #semanticClimate using material from the Intergovernmental Panel on Climate Change (IPCC/AR6) Working Groups (WG1, 2, 3).
Lucas Pretti: So if I ask something unrelated, the chatbot won’t answer?
Udita Agarwal: Exactly. Since that information is not present in our knowledge base, the chatbot will respond that it doesn’t know. It’s a climate chatbot, so it should answer climate-related questions, not general questions outside its scope.
Lucas Pretti: This has to do with RAG (Retrieval-Augmented Generation), right? This is the approach you used for building the chatbot. For people who don’t come from computer science, what is it?
Renu Kumari: The main idea behind RAG is that the chatbot retrieves information from a predefined set of resources rather than searching the entire internet.
Imagine you have two books. Instead of sending a language model to search the entire web, you tell it: ‘Only use these books.’ The retrieval component finds the relevant information inside those sources, and then the generation component turns that information into a natural-language answer.
Lucas Pretti: Is RAG the strategy for avoiding ‘hallucinations’ (another complicated term because it dismisses as an accident what may in fact be the very nature of this technology)?
Renu Kumari: We think of hallucination as a situation where a model generates information despite not actually having access to it. Large language models are trained to generate answers, not necessarily accurate answers. If you ask them a question, they will often try to provide a response even when they are uncertain.
We didn’t want that behaviour. So our approach is simple: if the information exists in our knowledge base, the chatbot can answer. If it doesn’t exist, it should say, “I don’t know.”
For example, if our sources don’t contain information about a political figure or some unrelated topic, the chatbot won’t attempt to answer. It only works with the information available in its curated climate knowledge base.
Lucas Pretti: Interesting. Internally at OKFN, we’ve sometimes been calling this informally as ‘AI responses with certificate of origin’, as a way to emphasise traceability. The guardrails that #semanticClimate talks about are basically these fixed, trusted resources?
Udita Agarwal: Exactly. The guardrails are the trusted resources that the chatbot is allowed to use. They act as hard boundaries that keep the system grounded in reliable information.
Lucas Pretti: Tell us how you index information semantically.
Udita Agarwal: Let me use an analogy. Suppose you want to find information about a specific person in a book. You wouldn’t read all 300 pages from beginning to end. You would go to the table of contents, find the relevant section, and then locate the information you need.
Computers need a similar structure. If we simply provide an entire book to the system and ask a question, it becomes very difficult and inefficient to find the relevant information. The model may miss important details, produce poor answers, or consume unnecessary computational resources.
So we divide the book into smaller pieces, or ‘chunks’. We first break it into sections and chapters, then into smaller units such as paragraphs.
When a user asks a question like “What is the greenhouse effect?”, the system searches the knowledge base and retrieves only the paragraph or paragraphs that contain the relevant information. Those chunks are then passed to the language model, which generates a readable answer.
Without ‘chunking’, we would be asking the model to process entire documents every time. That would be inefficient and much less reliable.
Lucas Pretti: When you say ‘knowledge base,’ what does that actually look like? Where does it live?
Renu Kumari: We use ChromaDB as our vector database. The Climate Academy book is first converted into a format that can be processed, then chunked, and those chunks are stored in the database.
The same process applies to the glossary and other resources.
Lucas Pretti: So if you wanted to add new sources in the future, you could expand the knowledge base? Or apply the same architecture for other domains, such as public data or education?
Renu Kumari: Yes, absolutely. We can add new resources and create additional databases that connect to the same system. That’s something we’re already exploring.
The overall architecture is domain-agnostic. You can apply the same process to research papers, government reports, educational materials, or other document collections. The main challenge is testing and selecting the most suitable large language model for the specific use case.
Lucas Pretti: What language model are you currently using?
Udita Agarwal: We are using Llama 3.1 8B. Selecting the model was actually one of the major technical challenges. Different models require different resources, and access to APIs and computing infrastructure is often limited. We wanted a model that was efficient while still providing reliable answers for our use case.
Lucas Pretti: The chatbot supports multiple languages. How does that work?
Udita Agarwal: The multilingual capability comes from the language model.
When a user submits a question in Hindi, Spanish, German, or another supported language, the system first converts that query into a representation that can be searched against the knowledge base. Once the relevant information is retrieved, the language model generates the answer in the user’s selected language.
The underlying climate knowledge remains the same; only the language of interaction changes.
Lucas Pretti: What was the biggest challenge during the project?
Renu Kumari: One challenge was infrastructure: obtaining servers, hosting the system, and managing backend resources. Another was access to language model APIs. The AI ecosystem is becoming increasingly closed, and many services have strict limits or costs associated with them.
Finding the right model, obtaining the necessary tokens, and ensuring that everything worked reliably required significant effort.
Lucas Pretti: Is there anything else you’d like readers to know?
Renu Kumari: This is still an early version of the chatbot. We see it as a starting point rather than a finished product. The most valuable thing now is feedback. We want people to use it, test it, and tell us what works and what can be improved. That’s how we’ll continue making it better.
This project has been a collaborative team effort. Aleena Harold Peter and Shreya Bhat designed and developed the frontend interface, Udita Agarwal designed and implemented the Retrieval-Augmented Generation (RAG) architecture, Ali Mehdi Naqvi developed the backend infrastructure and Renu Kumari served as the Program Manager, coordinating the project and managing its documentation.
The project also benefited from the contributions of team members who studied and structured Climate Academy’s materials to support the knowledge base. Testing was carried out by several contributors, including high school students, whose feedback helped improve the system.
We would like to sincerely thank Saarthak Patwal, Parth Kadam, Samyukta Nair, Bui Doan Khanh Linh, and the entire #semanticClimate team, especially Prof. Peter Murray-Rust and Dr. Gitanjali Yadav for their valuable contributions and support throughout the project.
About
Open Knowledge’s AI Learning Labs is an initiative that aims to experiment with AI, translate knowledge from social sector organisations around the world, and produce public, multilingual AI-literacy resources tailored for organisations addressing similar issues elsewhere.
Together, we will catalyse learning and develop replicable methods to help organisations build AI skills, use AI responsibly, and develop their own AI projects. All resources will be openly available at School of Data.
Join the conversation:
- Open Knowledge Forum
- School of Data community
- Contact our team: info@okfn.org
This project has been made possible thanks to the generous support of the Patrick J. McGovern Foundation (PJMF). We are grateful for our ongoing partnership in promoting digital literacy and investing in AI for the public good. Learn more about its funding programmes here.