Enhancing the performance of your ChatLab chatbot involves fine-tuning its settings and training methodologies. Start by training your chatbot, then select it and use Overview to ask a few real customer questions. Record the expected answers, change one thing at a time, and repeat the same questions in fresh conversations.
Custom Role Instructions for Fine-Tuning Responses
Custom Role Instructions allow you to define specific guidelines for your chatbot's behavior and tone, tailoring it to meet the needs of your audience. Here's how to implement and utilize this feature effectively:
-
Define the Bot's Role: Clearly specify the role the chatbot should assume. For example, "customer support agent," "sales assistant," or "technical advisor."
-
Set Behavioral Guidelines: Include instructions on tone, language, and response style. For example:
- Tone: Friendly, professional, or formal.
- Language: Use simple explanations for non-technical users or technical jargon for expert audiences.
-
Incorporate Context-Specific Details: Add instructions related to your business or domain. For example:
- "If a user asks about delivery options, prioritize mentioning free shipping if available."
- "When discussing pricing, always include details about bulk discounts if applicable."
-
Test and Iterate: Test how the chatbot responds with the new role instructions, refine them based on user interactions, and adjust as needed to achieve the desired response style and accuracy.
Configure this in Settings > Role & Behavior. Learn more about role and behavior settings.
Upgrade to a Stronger AI Model
Compare the models offered in the current selector on the same questions. Check whether the answer is accurate, follows your instructions and handles follow-up questions. Model availability and credit weights change, so use the displayed options and message-credit guide rather than a fixed model-price list.
Change your model in Settings > Model & Advanced. Learn more about selecting AI models.
Optimize Knowledge Base
Use the Knowledge Base Optimizer to see which fragments match a particular question. It tests retrieval for a standalone question, not historical source popularity. Read the returned content and compare it with your expected answer.
Select your chatbot and open Training > Knowledge Base Optimizer. Learn more about using the Knowledge Base Optimizer.
Additionally, reducing the creativity of the model can make responses more precise and on-topic. Lower creativity can reduce variation, but cannot correct missing information or guarantee factual accuracy. Availability depends on the selected model. You can adjust this in Settings > Model & Advanced using the Creativity slider. Learn more about conversation settings.
Context Expansion
Consider expanding context when relevant information does not fit. A larger context allows more material to fit, but does not guarantee a better answer. It also increases message-credit usage, and larger sizes require the corresponding account feature.
ChatLab offers three different context sizes: 8,000, 16,000, or 32,000 tokens, directly impacting the chatbot's ability to retain and utilize information. Choosing the appropriate context size is a balance between desired response quality and operational budget.
Configure context size in Settings > Model & Advanced. Learn more about extending chat context size.
Focused Training - Less is More
One of the most effective ways to improve response quality is to reduce irrelevant content in your training data. ChatLab uses RAG (Retrieval-Augmented Generation) architecture, which searches for the most relevant chunks of information when answering questions. Too much generic or repetitive content creates noise that can confuse the retrieval process.
Why focused training helps:
- RAG scores chunks of content for relevance - irrelevant content dilutes search quality
- Repetitive elements (menus, footers, headers) on every page flood the knowledge base with noise
- The AI may retrieve low-value chunks instead of specific, helpful answers
How to focus your training:
- Use sitemap scanning instead of full website crawls
- Exclude repetitive page elements like headers and footers
- Filter out irrelevant URL sections (news, images, press releases)
- Focus on valuable content: product pages, FAQs, support articles, policy documents
For detailed instructions on optimizing your website scans, see how to reduce training characters when scanning a website.
Preferred Format for Training Files
When training your chatbot, choose a file format that keeps related facts together. Text is useful for narrative explanations, while spreadsheets and CSVs can work well for tabular records:
-
Text files allow you to include contextually associated terms.
-
They ensure flexibility in structuring information.
For example, when adding a price table for products, include associated terms:
- Price Table Terms: "cost", "pricing", "rates", "fees", "charges", "quote"
Test whether queries like "What are your charges?" or "Show me the cost details" retrieve the right source and produce the expected answer.
Add Corrections from Chatlogs
Review your chatbot's conversation history and add corrections when you spot incorrect or suboptimal responses. This teaches the chatbot from its mistakes and improves future answers to similar questions.
Go to the Chatlogs tab, find conversations with issues, and click the edit icon next to any bot response to add a correction. Learn more about adding and editing corrections.