Expert Support case study: Bolstering a RAG app with LLM-as-a-Judge

This is a guest blog post authored by Digital Green. Digital green is participating in a CGIAR-led collaboration to bring agricultural support to smallholder farmers.

There are an estimated 500 million smallholder farmers globally: they play a critical role in global food security. Timely access to accurate information is essential for these farmers to make informed decisions and improve their yields.

An “agricultural extension service” offers technical advice on agriculture to farmers, and also supplies them with the necessary inputs and services to support their agricultural production.
Agriculture extension agents are 300K in India alone, they provide necessary information about improved agriculture practice and help in decision making for the smallholder farmers.

But although their number is impressive, extension workers are not in large enough numbers to cope with all the demand: they interact with farmers at typically in the ratio of 1:1000. Reaching the agriculture extension workers and farmers through partnership and technology remains the key.

Enter project GAIA, a collaborative initiative pioneered by CGIAR.
It brought together Hugging Face as mentor through the Expert Support program, and Digital Green as project partner.

GAIA has a lofty goal to bring years of agriculture knowledge in the form of research papers meticulously maintained in GARDIAN portal in the hands of the farmers. There are close to 46000 research papers and reports that have agricultural knowledge globally carried over multiple decades across different crops.

Digital Green immediately saw the potential of developing intelligent chatbots powered by Retrieval-Augmented Generation (RAG) on approved, curated information. Thus they decided to develop Farmer.chat, a chatbot that leverages the capabilities of large language models (LLMs) to deliver personalized and reliable agricultural advice to the farmers and front line extension workers.

Creating such a chatbot for a huge variety of languages, geographies, crops, and use cases, is a gigantic challenge: information disseminated has to be contextual to the local level details about the farm, in the language and tone that farmers can understand and accurate (grounded in trustworthy sources) for farmers to act on it. To evaluate the performance of the system, the CGIAR team and HF expert collaborated to set up a strong evaluation suite, in the form of an LLM-as-a-judge system.

Let’s take a look at how they tackled this challenge!

System architecture

The full system uses many components in order to provide chatbot answers grounded in several tools and external knowledge. It has several key elements:

• Knowledge base:
  • Preprocessing: The first step was to ingest the pdf documents into the Farmer.chat pipeline with the help of APIs maintained by Scio. The in the knowledge base, topics were auto categorized for relevant geographic areas and semantically grouped together.
  • Semantic chunking: the organized files with metadata are processed with sentences similar in meaning grouped together in text chunks. The function uses small-text embedding currently for cosine similarity
  • Conversion into VectorDB format: each text chunk is converted into vector representation using an embedding model using which the vector representation is stored in QdrantDB.
• RAG pipeline: It is what ensures that the information delivered is grounded in the content and not outside. It consists in two parts:
  • Information retrieval: Searching the knowledge base for relevant information that matches the user’s query. This involves calling the vector database API created in knowledge base builder to get necessary text chunks.
  • Generation: Using the retrieved information in the text chunks and the user query, the generator calls LLM and generates a human-like response that addresses the user’s needs.
• User-facing Agent: The planning agent leverages GPT-4o under the hood.
  • Its task is to:
    • Understand the user intent
    • Based on the user intent and tools description decide what more information is required
    • Ask that information from the user till the ask is clear
    • Once the ask is clear, call the execution agent
    • Get the response form the execution agent and generate response
  • The agent runs a ReAct based prompt to think in step by step manner and call the respective tools and analyze the responses. Then it can leverage its tools to answer: Currently, the agent uses the following set of tools:
    • Converse more
    • RAG QA endpoint
    • Video retrieval endpoint
    • Weather endpoint
    • Crop table

Now this system has many moving parts, and each part has a radical impart on some aspects of performance. So we need to carefully run performance evaluation.

In the last one year, the usage of Farmer.chat has grown to service more than 20k farmers handling over 340k queries. How can we evaluate the performance of the system at this scale?

During weekly brainstorming sessions, Hugging Face hinted to LLM as a judge and provided a link to their notebook LLM as a Judge. This resource was discussed in detail, and what followed became a practice that has helped navigating Farmer.chat’s development.

The Power of LLMs-as-Judges

Farmer.Chat employs a sophisticated Retrieval-Augmented Generation (RAG) pipeline to deliver accurate and relevant information to farmers that is grounded in the knowledge base. The RAG pipeline uses an LLM to retrieve information from a vast knowledge base and then generate a concise and informative response.

But how do we measure the effectiveness of this pipeline?

The difficulty here is that there is no deterministic metric that one could use to rate the quality of an answer, its conciseness, its precision...

That is where LLM-as-a-judge technique steps in. The idea is simple: ask an LLM to rate the output on any metric. The immense advantage is that the metric can be anything: LLM-as-a-Judge is extremely versatile.

For example, you can use it to evaluate the clarity of a prompt as follows:

You will be given a user input about agriculture, and your task is to score it on various aspects.
Think step by step and rate the user input on all three following criteria and give a score for each:
1) The intent and ask is clear.
2) The topic is well-specified.
3) The target entity is well-specified, as well as its attributes, for instance "disease resistant" or "high yield".
You should give your scores on an integer scale of 1 to 3, 1 being the worst and 3 the best score.

After creating a score for each three, take the average and round it off to the nearest integer which becomes the final score.  

Example:
User input: "tell the benefits of batian coffee variety"
Criterion 1: scores 3, as the intent is clear (about knowing about batian variety of coffee) and the ask is clear (want to summarize the benefits).  
Criterion 2: scores 3, the topic is well specified (coffee varieties)   
Criterion 3: scores 2, as the entity is clear (batian variety) but not the attributes.

As mentioned in this article that we referred to earlier, the key to use LLM-as-a-judge is to clearly define the task, the criteria and the integer rating scale.

The research team behind Farmer.Chat leverages the capabilities of LLMs to evaluate several crucial metrics:

• Prompt Clarity: This metric evaluates how well users can articulate their questions. LLMs are trained to assess the clarity of user intent, topic specificity, and entity-attribute identification, providing insights into how effectively users can communicate their needs.
• Question Type: This metric classifies user questions into different categories based on their cognitive complexity. LLMs analyze the user's query and assign it to one of six categories, such as "remember," "understand," "apply," "analyze," "evaluate," and "create," helping us understand the cognitive demands of user interactions.
• Answered Queries: This metric tracks the percentage of questions answered by the chatbot, providing insights into the breadth of the knowledge base and the platform's ability to address a wide range of queries.
• RAG Accuracy: This metric assesses the faithfulness and relevance of the information retrieved by the RAG pipeline. The LLM acts as a judge, comparing the retrieved information to the user's query and evaluating whether the response is accurate and relevant.

It empowers us to go beyond simply measuring how many questions a chatbot can answer or how quickly it responds. Instead, we can delve deeper into the quality of the responses and understand the user experience in a more nuanced way.

For RAG accuracy we use LLM-as-a-judge to evaluate on a binary scale: zero or one. But the way the task is broken down leads to a well established process that comes up with a score that we tested with human evaluators on roughly 360 questions: LLM answers are found to actually do a great job and have high correlation with human evaluations!

Here is the prompt, which was inspired from the RAGAS library.

You are a natural language inference engine. You will be presented with a set of factual statements and context. You are supposed to analyze if each statement is factually correct given the context. You can come up with the scores of 'Yes' (1) and 'No' (0) as verdict.

Use the following rules:  
If the statement can be derived from the context, give a score of 1.
If there is no statement and there is no context, give a score of 1.
If the statement can’t be derived from the context, give a score of 0.
If there is no context but there is a statement, give a score of 0.

#### Input :

Context : {context}

Statements : {statements}

The context variable above is the input chunks given for generating the answers while statements are the atomic factual statements generated by another LLM call.

This was a very important step as it enables evaluation at scale which is important when dealing with large numbers of documents and queries. The LLM-as-a-judge at core leads to metrics that act as a compass navigating the various options available for our AI pipeline.

Results: benchmarking LLMs for RAG

We created a sample dataset of > 700 user queries randomized across different value chains (crops) and date (months). While this upgrade itself had 11 different versions that evaluated using RAG accuracy and percentage answered, the same approach was used to measure performance of the leading LLMs without any prompt changes in each LLM call. For this experiment, we selected GPT-4-Turbo by OpenAI, Gemini-1.5 in Pro and Flash versions, and Llama-3-70B-Instruct.

What we see is that amongst the four models, the highest level of factually correct answers (“Faithful” column) is obtained with Gemini-1.5-pro, followed very closely by Gemini-1.5-Flash and GPT-4-turbo.

What we found was that purely on the basis of faithfulness, Gemini-1.5-Pro beats the other models. But if we also take into account which percentage of questions the model accepted to answer, Llama-3-70B and Gemini-1.5-Flash perform better.

In the end, we picked Gemini-1.5-Flash due to the superior trade-off of a low percentage of unanswered questions and very high faithfulness.

Conclusion

By leveraging LLMs as judges, we gain a deeper understanding of user behavior and the effectiveness of AI-powered tools in the agricultural context. This data-driven approach is crucial for:

• Improving user experience: By identifying areas where users struggle to articulate their needs or where the RAG pipeline is not performing as expected, we can improve the design and functionality of the platform.
• Optimizing the knowledge base: The analysis of unanswered queries helps us identify gaps in the knowledge base and prioritize content development.
• Selecting the right LLMs: By benchmarking different LLMs on key metrics, we can make informed decisions about which models are best suited for specific tasks and contexts.

The ability of LLMs to act as judges in evaluating the performance of AI systems is a game-changer. It allows us to measure the impact of these systems in a more objective and data-driven way, ultimately leading to the development of more robust, effective, and user-friendly AI tools for agriculture.

In the span of over a year, we have continuously evolved our product. In this small timeframe we have been able to:

• Reach more than 20k farmers
• Answer > 340k questions
• Serve > 6 languages, for 50 value chain crops
• Maintain close to zero biases or toxic responses

The results were published recently in this scientific article, focusing on the quantitative study of user research.

If you are interested in the Hugging Face Expert Support program for your company, don't hesitate to contact us here - our sales team will get in touch to discuss your needs!