The world runs on a lot of what I call qualified blind trust. That is, we don't know the people behind most of our daily systems, but we trust them to cook our food, hold our money, give us prescriptions, or even fly our planes. We trust them not because they told us we should, but because we trust the training, testing, and frequent evaluation and certification needed to maintain their roles. So when someone asks why you should trust another AI chatbot like Suwali – well, you shouldn't because we asked you to.
Instead, you should ask, "do I trust the process that the chatbot has undergone?"
Since trust is earned by trusting the process, we want to show you, in this blog, several of the many ways we evaluated Suwali before deciding it could work for you. There are three key components for evaluating Suwali and similar chatbots: a golden dataset, automated evaluations, and human evaluations.
The golden dataset
The golden dataset is the basis of our evaluations of Suwali. They are called golden because they represent the standard by which we measure Suwali responses. The datasets were developed by our regional Program Managers, with several having input from regional partners who are all native or fluent speakers of Hindi, Arabic, Brazilian Portuguese, English, and French. These datasets consist of real questions across a diverse range of topics, from local pollution laws to global AI safety, and from reproductive health to ongoing wars. All questions we expect a Suwali user to ask. The questions are paired with a human-written answer, which we call the golden answer, formed using selected human-written sources. The same questions are posed to Suwali, and we evaluated each of its responses with a diverse set of tests.
Automated evaluations
I like to think of Suwali as a student who answers questions by learning from reading and the knowledge of trusted users. Students undergo regular testing to ensure they are meeting expectations and we do the same. The tests consist of both automated and human-in-the-loop evaluations that cover all aspects of Suwali's operations. These fall into two areas we wanted to make sure Suwali got right: retrieval and generation.
Retrieval is the step where the student goes to the library to search for the right books, and Suwali is like having a dedicated librarian who helps the student "retrieve" the right books for the question among the hundreds or thousands of books. We evaluated the retrieval ability by knowing that there are exactly three books in the library that answer our questions. The librarian might also bring back other books they think might be helpful, so we can check the top 5 recommended books. If they bring back all 3 books within their top 5, they get 100%; if only 2 are brought back, that's 66% (⅔), which is called recall@5 in technical jargon. This is exactly how we evaluated whether Suwali is finding the right information to base its answers on, and we are proud to say that Suwali achieves 83% across all datasets.
With the books that the librarian brought back, the student can now start answering their essay question. This step, where a large amount of information is synthesized, is called generation in Suwali. Evaluating the generated response is tricky because two different students can present the same information in very different ways. If asked the topic question: "Describe how plants get their energy?" One may write "plants get energy from sunlight" while the other answers "plants convert sunlight, water, and carbon dioxide into carbohydrates and oxygen". Neither is wrong, as they offer two correct but different answers with varying degrees of facts. LLM answers also have no single right answer, and the best evaluations end up being "it depends," which means we don't simply compare Suwali's response to the golden answer. We evaluate generation more holistically using a range of automated and human-in-the-loop evaluations.
To make a subjective task slightly repeatable, we asked the authors of the golden dataset to include a list of facts that the answer must include to be considered correct. Similar to how an essay on the same question can earn the same points by presenting the same information differently. We call this the fact completeness score, which we achieved 84% before we launched Suwali.
We go one step further to test that Suwali didn't answer solely from its internal knowledge by assessing the grounding of its reply by comparing it with the retrieved documents, like how a teacher may judge how well responses are formed based on an assigned reading list. We do this using the eponymous groundedness metric, which roughly corresponds to a score based on the percentage of retrieved sources in which the answer appears. This metric achieved only 70% before the launch, giving us room to improve the answer.
Human evaluations
The last step of the evaluation is the ultimate one, by an all-knowing oracle, from the perspective of Suwali, the original golden dataset authors. We invited the authors to blind judge four versions of each answer generated by four different models. They voted on the answers, and we deployed the model that received the most votes. Their votes were based on a number of assessments: whether the tone suited the context (a newsroom to be more factual and formal, while a health organization to be more empathetic), whether the answer fits the natural use of the language, and most importantly, which one simply read better while answering the question. This evaluation is admittedly subjective, but it was important to ensure that the answers meet users' needs from regional and linguistic perspectives.
The evaluation of Suwali was not a one-off task but a recurring process we run to ensure we could track Suwali's performance. We are constantly expanding our golden datasets with new topic areas and more difficult questions. We are also constantly adding new metrics to assess qualities that our current metrics miss, such as safety, language adherence, and how well Suwali refuses to answer. The evaluation of Suwali is a rigorous, repeatable assessment we have built to ensure that the version you are using meets the standard we expect.
