What to do if AI Search and AI Q&A are inaccurate? How to retrain AI results?

Reasons and Solutions for Inaccurate AI Responses

Content Updated During a Conversation

Reason: If new content is added or updated during an ongoing conversation, the AI may not have learned it yet. Additionally, since HelpLook supports multi-turn conversations, the AI incorporates context from previous exchanges to interpret questions, which might affect accuracy.

Solution: Refresh the page and re-enter the question, or click the "+" icon in the upper-right corner of the dialog box to start a new conversation.

Incomplete Data

Reason: AI generates responses based on the data it has received. If the data is insufficient or of low quality, the AI’s responses may be limited.

Solution: When testing AI’s performance, ensure all relevant materials are uploaded. You can also check the AI Q&A  to view detailed responses and source data snippets to identify any missing content.

Excessive Document Length

Reason: Large language models (LLMs) have a token limit, so they summarize answers based on the top six most relevant fragments.

Solution:

Split lengthy documents into multiple smaller ones.

For inaccurate responses, draft a standard reply, and retrain the AI using that reply. You can review the AI’s responses and source data in the AI Q&A to ensure relevant data is included.

Draft or Pending Articles

Reason: Ensure the article is in a published state. Only published articles are transmitted to the AI model for training. Draft articles will not be included in AI training.

 "·" indicates the draft state.

Impact of Different AI Models

Reason:

Different AI models vary in terms of completeness, accuracy, and other aspects. 

Additionally, LLMs have token limits, so only the top six matching fragments are used for intelligent summaries, making them unsuitable for site-wide aggregation queries.

Solution:

Test different AI models to compare results and select the one that performs best.

Additionally, LLMs have token limits, so only the top six matching fragments are used for intelligent summaries, making them unsuitable for site-wide aggregation queries.

For inaccurate responses or site-wide aggregation queries, draft a standard reply and retrain the AI with that content.

Text Match Rate

Reason: A lower text match rate results in more generalized AI responses, while a higher match rate yields more precise answers. However, if the question has low relevance to the content, the AI may not provide an answer.

Solution: Use the default value of 60 for the text match rate.

Impact of AI Prompts

Reason: Customizing Prompts allows you to train the AI’s role, set default responses, and generate more personalized outputs, such as specific text styles, tones, or themes.

Solution:

Click Setup -> Site Settings -> Search, and find Custom Prompt, and configure the Prompt for AI Search.

In AI Widget -> Settings, locate Custom Prompt to set up the Prompt for the AI Q&A bot.