Talkative GenAI Chatbot - AI Knowledgebases

Talkative GenAI Chatbot - answer questions using an LLM and your own dataset!

Talkative now supports Generative AI responses in the Talkative chatbot builder.

This guide is split into three key sections:

1. How it works

2. Creating Knowledgebases

3. Using the Knowledgebase in the chatbot

This forms part of the Talkative chatbot builder at guide. Read more about this here: Talkative Chatbot Builder

1. How it works

Talkative's chatbot supports integration with an AI knowledge base.

The key concept of the generative AI chatbot is using a Large Language Model (OpenAI) to respond to a customer's question, based on a pre-approved data set (e.g. your website).

The interaction flow works in the following order:

1. A customer starts a chat interaction with your chatbot, and sends in a message.

2. If the chatbot flow uses an AI Knowledgebase lookup to answer that particular question, the message will be sent to the LLM.

3. The LLM is prompted to respond as if it were a customer service agent, answering only based on the dataset you've provided.

4. If the LLM cannot answer the question, it will return a system message to your chatbot, from where you can make the next action in the chatbot flow (i.e. route to an agent).

5. If the LLM can answer the question, it will send the response back to the Talkative system, where it is inserted into the chatbot response.

No data is stored with the LLM. No identifying metadata is sent to the LLM - all pre-chat forms data, company name, queue name, agent name etc. are all stripped out before the request is sent over. The LLM is only aware of the dataset provided, the prompt, and the conversation thread including the message it needs to respond to.

2. Creating Knowledgebases

Pre-requisites

• Talkative' GenAI feature is currently available as a closed beta. Please contact Talkative if you would like to try this feature. 
• You must first have a Talkative chatbot set up which you can follow a guide for here: Talkative Chatbot Builder

1. Navigate to "AI Knowledge Bases" in the settings and press Create.

2. Knowledge Base supports CSV and TXT files, as well as websites and webpages.

3. File based KBs are limited to 10MB. Webpage-based are limited to 100 URLs.

4. URL based KBs are currently limited to 100 pages per Knowledgebase. This limit will change/increase in future.

5. You can scan your website to generate up to 100 URLs at once.  If you include anything past the url, only pages that have that will be included, ie if you do https://example.com/help only pages that have /help in the url will be part of the list

6. Once your KB is created, Account Holders in your Talkative system will receive an email when the KB is finished uploading. You can also manually check this by clicking into the KB's advanced options settings. Uploads typically take 1-2 minutes maximum.

7. [Optional] You can choose to add a supplementary prompt to your KB lookup under advanced section. We'd recommend leaving this empty by default. Max 1,000 characters.

That's it, you're ready to go!

3. Using the Knowledgebase in the chatbot

1. Navigate to your Talkative chatbot (assuming this is set up already, but you can create one from scratch if needed).

2. Decide where in the chatbot flow you want to use the AI lookup. There are a few approaches:

i) Having a lookup in a specific flow. This can be useful if you want 

ii) Using GenAI lookups in the fallback.

iii) Using GenAI response as your initial response to the customer.

We'd advise thorough testing of your chatbot if you are using option iii) in particular. 

Let's show how you'd do each of these options:

3. i) A lookup in a specific flow 

In this case, we select the exit approach of the node to be "Look up previous message in AI knowledge base"

So your flow will look like this:

Let's preview this to see how it works. First, the chatbot asks the user to input their question:

Once the question has been received, the chatbot will do an AI lookup, querying the question against the KB specified. 

In this case you can see the data was matched, and we include the response here:

3. ii) Using GenAI lookups in the fallback.

In this case we're using the AI lookup as the fallback node.

Here we're initially asking the customer to go down our path of suggestion chips, where we've got scripted responses. But if the customer veers off from our decision tree, they will hit the fallback node.

In this initial conversation, we go down the "expected" path we want the user to take.

But in this second conversation, our question is not covered by our pre-written responses.

The fallback node is triggered because of this. And the fallback node triggers the AI Knowledgebase lookup.

3. iii) Using GenAI response as your initial response to the customer.

Here the initial message is set to go to the AI Knowledge lookup:

4. FAQs

Q: Will this answer questions that are not relevant to my business?

A: No, the AI will only respond if the answer to the question is within your dataset / Knowledgebase.

Q: My knowledgebase hasn't uploaded.

A: If your KB is within the data size limit and you still have issues, please contact Talkative.

Q: How much does it cost?

A: During the beta, there is no cost for using the AI features. Pricing will be available later in Q1 2024 for all Talkative customers to use.

Q: Does it matter how my data is structured?
A: LLMs work well with unstructured data, but FAQ formats tend to work the best. The higher quality of data, the better your results will be. 

Q: How many URLs can I use?

A: Currently just 100, although this limit is getting increased soon.

Q: How many Knowledgebases can I use?

A: There is currently no limit.

Q: What benefits can I expect to see from this?

A: This really depends on what sort of customer questions you are receiving. The main benefits are that the AI is going to be better at responding to questions than traditional chatbot, allowing it to answer a higher percentage of questions without human assistance.

Q: How well does it work on websites - dropdowns?
A: The KB import tool works well on most websites, however there will inevitably a small percentage of websites that it doesn't work well on. Please contact Talkative if you require support with importing

Q: What's the difference between fallback and data not matched nodes?

A: Fallbacks are your global Talkative Chatbot node for when a user's message input is not matched to a flow. The "item/data not matched" nodes are for when the AI lookup tool was used, but the data/answer was not found in the dataset/KB.

Data

Only the message transcript is sent to OpenAI, pre-chat form data and other interaction metadata is not used. OpenAI does not store any messages or responses. Please familiarise yourself with OpenAI's privacy policies before using the system. https://openai.com/policies/api-data-usage-policies

OpenAI will be added as a data sub-processor if you use this feature. 

5. Roadmap

Talkative has a number of enhancements planned for the GenAI feature. Please contact us to learn more!