How accurate Wallu’s responses are and how well it performs depends on the information it has access to and how you’re using it. Here we will cover some tips and tricks and common pitfalls to avoid to get the most out of Wallu.

FAQs vs Documents

Wallu has two main sources of information: FAQs and Documents. FAQs are manually curated and are the most reliable source of information. Documents work more like black boxes. Wallu can automatically extract information from them and use it to answer questions. There is less work to manage documents but there is always the possibility Wallu misinterprets something and provides an incorrect answer.

Tips & tricks for FAQs

• Prefer full sentences. Instead of, “request refund” or “how refund”, just use “how to request a refund?”. The bot is trained with examples, not keywords.
• For FAQs, we recommend using the “send exactly as written” as it means the AI will not change the answer and it will be as accurate as the FAQ.
• It’s usually good idea to in referene to the question in the answer. E.g., “Yes, we do offer refunds.” Instead of just “Yes.”. This way if the bot incorrectly answers a different question with this answer, the user sees that the answer is not relevant to their question.
• Even though the AI is good with synonyms, it’s beneficial to list multiple different questions. For example, “do you offer refunds?” and “where do I request a refund?”. However, no need to list too similar questions like “where can I track my order?” and “how can I track the order?”.
• The AI does not know your community specific terms. For example, Wallu most likely does not know that BA in “how to apply to BA?” is a synonym for “British Airways”.

Common Pitfalls with Documents

Incomplete sentences (lists/tables/links instead of paragraphs)

AI based search works better when there are complete sentences or paragraphs for the question.

Recommendation: Ideally the document should be “one paragraph per topic”. If you have a list of items, try to make each item a complete sentence. This way the AI will pick the most relevant paragraph to the user’s questions.

For example, instead of a document like this:

Plans:
Premium $5/m
Enterprise $10/m

Social media:
Youtube: https://youtube.com/mybusiness
Website: https://example.com/mybusiness

Reset password

Visit here: https://example.com/reset-password


Change billing:

https://example.com/billing

Prefer a document more like this:

We offer two plans: Premium costs $5 per month. Enterprise costs $10 per month per user.

You can find mybusiness on social media:
mybusiness Youtube channel: https://youtube.com/mybusiness
mybusiness website: https://example.com/mybusiness

You can reset your password here: https://example.com/reset-password

You can change your billing information, cancel your plan or change your plan here: https://example.com/billing

This way the AI can pick the most relevant paragraph to the user’s question. For example, if the user asks “how do I change my plan?”, the AI will more easily find the correct paragraph now.

TIP: Consider using ChatGPT or other AI based tools to generate text from lists and tables to make them more readable for Wallu.

Garbage in, garbage out

If you have a document with a lot of outdated information, you cannot expect Wallu to provide accurate answers. More documents does not necessarily mean better results. Also, even though importing your entire website is very easy way to get started, the format, information and layout may not be as optimal for AI.

Recommendation: Try to focus on quality over quantity. If needed, copy-paste your documents and then edit them to make them more readable for AI (see above section for more help). You can also consider using more documents in #ai-support channels where user’s know the answers are obviously AI generated and then use the highest quality documents & FAQs on other channels for automatic answers.

Review content regularly

Wallu offers the optimizer tool to help you identify information that need to be updated or may be misleading (either Wallu does not properly understand it or it is even misleading to users).

Optimize now: https://panel.wallubot.com/optimizer

Choose the AI model

Additionally to Wallu’s custom AI models, it uses models provided by OpenAI.

Lite AI (GPT-4o mini):

• Cheapest model
• Least accurate answers
• Ok for general topics, but not recommended for business critical information
• Do not use if inaccurate answers are a problem
• Mainly meant for testing and general topics (e.g., creative writing, general knowledge)

Advanced AI (GPT-4o):

• Most accurate answers (less likely to provide incorrect answers, make up things and misunderstand users or documentation)
• Use if you want the most accurate answers about your product/service/business
• Recommended for most use cases
• Uses more credits (about 10x)

Recommendation: Please prefer Advanced AI (GPT-4) option if you want the most accurate answers. The Lite AI is mainly meant for testing, more general topics (e.g., not specific to your product/service/business) or if you are on a very tight budget and do not care that much about accuracy.

Choose how Wallu searches the documents

In the admin panel, you can choose how thoroughly Wallu should search the documents (“Document Search Level”). The options are: Standard, Enhanced, Advanced and Expert.

Generally lower levels are more accurate, use less credits but will not find answer to all questions. We recommend starting with the Advanced or Expert level. If you want to save credits, you can try lower levels if the bot does not find answers to all questions try the higher levels.

Standard uses the least credits as it only answers questions that are obviously answered in the documents which also makes it more accurate. However, it may not find answer to all questions.
Enhanced uses slightly more credits and answers more questions.
Advanced uses more credits to search more thoroughly and answers even more questions. However, it is more likely to provide incorrect answers as it will try to answer more questions even if it’s not so obvious whether the document contains the answer.
Expert Uses most credits to perform extensive search and will try to answer all questions that can be answered based on the documentation. If you have chosen the Lite AI model, this is more likely misunderstand the documentation.

Recommendation: Start with the Expert model. If you are running out of credits or the bot is misinterpreting the documentation, try lowering the levels.