How to improve accuracy

Using best practices and optimizing the AI Knowledgebase is key to accurate answers. While the training process is fully automated, achieving the best results requires human intervention to improve data sources quality and optimize prompts.

Data sources

Format

• Adding whole website
  When using your website as a data source it is necessary to only include relevant pages and use the correct CSS selector to filter out unnecessary data (menus, headers, footers…). AI Knowledge cannot a will not answer accurately when 50% of indexed data is not relevant or readable by human, not even AI.
  Always check the data we indexed from your website manually in AI Knowledgebase.

• Text-Based Content
  All information should be in text form. The content must remain understandable even if all images, videos, and other media are removed. Common problem with PDFs or websites.

• Tables/Charts
  When tables or charts are converted to text form, lot of information is lost due to loss of structure. Information needs to be in text, not in the structure, format or layout - imagine reading a table on braille display.

• Avoid Text in Image Format
  PDFs or documents with text, tables, or any other information saved as images should be converted into editable text. AI Knowledgebase cannot read image-based text.

Structure

• Avoid Redundant Explanations
  Ensure each topic is covered only once in the data sources to avoid conflicts. Common error is including multiple data source types (web page + FAQ document) while both include slightly different answer to one question (or the answer on web page is updated).

• Use FAQ Format When Possible
  Write content as question-and-answer pairs for clarity and easy retrieval. You can include multiple variants of questions. Consider recreating your knowledge base in our Manual QnA option in AI Knowledgebase to achieve best results.

• Visual Separation
  Ensure topics are visually distinct using headings, subheadings, or bullet points.

Copywriting

• Restate Answers
  Answers should provide full-sentence responses, not just “yes” or “no.” This allows the AI to reuse the answer effectively. Also include full context in every sentence rather than assuming it’s obvious from context/location. This is especially important when not using question-answer pairs.

• Repeat Key Information
  Don’t rely on context to make important details clear; repeat where necessary.

• Clarity and Simplicity
  Use straightforward sentences. Ambiguous phrasing can confuse AI, leading to inconsistent responses.

• Explain Abbreviations
  Always define abbreviations within each text block, even if they are used elsewhere in the document. Avoid internal jargon.

• Consistent Tone
  Maintain a uniform writing style.

While LLMs are usually able to process data even without adhering to all these rules, it will ensure more accurate answers in cases when user questions are not clear, contain grammatical or speech to text transcriptions errors.

We are also strictly limiting our models to answer only when the answer is in your data sources to prevent them from hallucinating - that’s why we need to be able to search in your data effectively.

Prompts

Using prompts is optional and often not necessary. Good approach is to start without prompts to optimize your data sources and then use prompts for fine tuning or setting details like tone of voice etc.

Main reply prompt

Main prompt used for generating the reply. Use short, straightforward instructions.

Examples:

• Answer informally, use feminine (languages with grammatical gender)

• Answer strictly in Czech language

AI Knowledge filter prompt

This prompt is used to filter out unrelated data before generating a final answer. AI Knowledge tries to find all possible relevant data in your data sources based on users question. Then it tries to narrow down the most accurate ones based on internal logic and this prompt.

AI Knowledge query

Prompt used for generating more variants of user questions. Your customers could ask vaguely or use different words than you are using in your data sources. We can expand this input question to search in your data sources more effectively.