ChatLLM Training for End Users

Learning Objectives

The learning objectives of this article are:

• To learn how to use ChatLLM effectively for end users
• To learn about the different options available in the UI

Basic Navigation

The UI, after signing into the platform, should look similar to below:

Here are the important options:

(1): Use this button to upload your own documents. Most chatbots would already have access to some predefined documents based on the use case.

(2): Here you can find all the previous discussions you had with the chatbot

(3): Toggle between the different chatbots

(4): UI customisation and personal account details

Maximizing Chatbot Effectiveness: Tips for Better Responses

The quality of responses generated by AI systems is influenced by three main factors:

1. The Quality of the User's Input
2. The Quality of the Retrieval-Augmented Generation (RAG) System
3. The Quality of the Language Model (LLM)

The Quality of the User's Input

The clarity and specificity of user input are crucial. A well-crafted query not only helps the LLM provide better answers but also enhances the RAG system's performance. To maximise the effectiveness of AI systems, users should aim to be explicit and use straightforward instructions. The more relevant keywords you include, the better the system can understand and respond to your query.

• Example of a General Prompt: "How do I change oil?"
• Potential Issue: The document retriever might return documents related to various car models.
• Example of a Specific Prompt: "How do I change oil for my Audi Q3 2020?"
• Improvement: The document retriever is more likely to return accurate and relevant information.

Industry-Specific Examples

• Healthcare:
• Bad Prompt: "Tell me about diabetes."
• Good Prompt: "What are the latest treatment options for Type 2 diabetes in adults?"
• Finance:
• Bad Prompt: "Explain investment strategies."
• Good Prompt: "What are the best investment strategies for a risk-averse individual in 2024?"
• Technology:
• Bad Prompt: "How does cloud computing work?"
• Good Prompt: "What are the benefits of using AWS for small businesses in terms of cost and scalability?"

By following these guidelines and examples, users can significantly improve the quality of responses they receive from AI systems, leading to more accurate and useful outcomes.

The Quality of the RAG System

Improving the RAG system can significantly enhance response quality. Here are some backend strategies to fine-tune RAG:

• Increase Chunk Size: Documents are split into chunks, and adjusting the size can improve retrieval accuracy.
• Experiment with Different Embeddings: Testing various embeddings can lead to better document matching.
• Keyword re-ordering: Reordering based on keywords can optimise search results.

The Quality of the LLM

While the quality of the LLM is not something we can directly control, it's important to note that LLMs are continually improving. Abacus.AI is committed to supporting the latest advancements in LLM technology

[Optional] The RAG Process

At a high level, Retrieval-Augmented Generation (RAG) involves enhancing a large language model's (LLM) responses by retrieving relevant contextual information from a set of documents. This process integrates user input with pertinent data to generate more accurate and informative answers.

1. User Input: The process begins when a user submits a question or prompt.
2. Document Retrieval: Our Document Retriever searches for semantic similarities between the user's input and pre-existing document chunks. This step ensures that the most contextually relevant information is identified.
3. Contextual Integration: The retrieved document chunks are then provided to the LLM. This additional context helps the model understand the nuances of the user's query.
4. Response Generation: The LLM utilizes the contextual information to generate a comprehensive and accurate response.

In the Abacus.AI External Chat UI, users can view the document chunks that were retrieved during this process. Simply click on "sources" located just below the LLM's response to see the supporting information.

If the document retriever fails to surface the correct information, the LLM will struggle to answer the user's question accurately. Additionally, while the order of the document chunks has some importance, it has minimal impact on the LLM's ability to provide a correct answer. As long as the relevant document chunk is included, even if it appears last, the LLM can still generate an accurate response.