Question 1

Which file format should I use?

Accepted Answer

Use a plain text file (.txt) whenever possible. It is the most reliable format because what you write is exactly what the AI receives — no formatting gets lost or scrambled. PDF works well for simple, single-column documents, but complex layouts, tables, columns, or scanned PDFs often produce scrambled or incomplete text. DOCX works well for straightforward Word documents, but headers, footers, sidebars, and text boxes may be extracted in the wrong order or skipped. TXT is always extracted exactly as written and is recommended for all Q&A content. If your PDF or Word document is not performing well, export it as plain text and re-upload.

Question 2

Why does structure matter so much?

Accepted Answer

When your file is uploaded, the system splits the content into small chunks and generates an AI embedding for each chunk. When a caller asks a question, the AI finds the most semantically similar chunk and uses it to answer. The problem with unstructured text is that if a question and its answer get split into different chunks, the AI may find the question but not the answer, leading to vague or wrong responses. The solution is to use a Q&A format. The system automatically detects Q&A patterns and keeps each question and answer together as a single chunk — this is the single biggest improvement you can make to your knowledge base quality.

Question 3

What is the best format for a Q&A file?

Accepted Answer

Use the Question / Answer format with numbered entries. This is the most reliably detected pattern: "Question 1: What are your opening hours? Answer: We are open Monday to Friday from 9 AM to 5 PM. On Saturday we are open from 10 AM to 2 PM. We are closed on Sunday and public holidays." followed by "Question 2: How do I cancel my appointment? Answer: You can cancel by calling us at least 24 hours in advance, or by replying to your confirmation email." Each question and answer stays together as one unit so the AI always retrieves the full answer when it finds a matching question.

Question 4

What other formats does the system detect automatically?

Accepted Answer

The system recognises several Q&A patterns automatically. Option A (recommended, numbered): "Question 1: [question]? Answer: [answer]". Option B (Q shorthand): "Q: [question]?" followed by the answer on the next line. Option C (Markdown headings): "## [question]?" followed by the answer. Option D (numbered list with question marks): "1. [question]?" followed by the answer. All four options keep each question and answer together automatically. Option A (Question/Answer numbered) is the most explicit and recommended for new documents.

Question 5

What rules should I follow when writing the file?

Accepted Answer

Start each question on a new line (pattern detection looks for a line break before each new question). Include at least 2 questions (the system requires a minimum of 2 question-answer pairs to confirm Q&A format — single entries are better added as Quick Answers). One question per section (do not put multiple questions under one heading or answer block). Keep answers under 1,500 characters (very long answers get split anyway — if genuinely long, break into two separate questions). Be specific and complete (include prices, times, addresses, and conditions). No introduction or preamble needed (start directly with Question 1 — an intro paragraph becomes its own chunk with no question attached). Use plain language (write answers the way a helpful person would say them on the phone).

Question 6

What should I avoid?

Accepted Answer

Avoid free-flowing paragraphs without structure — if the system cannot detect a Q&A pattern, it splits content every 2,000 characters regardless of where questions and answers fall, which is the most common cause of poor AI answers. Avoid PDFs with multiple columns, tables, or scanned images — text extraction reads left to right across the full page width, scrambling multi-column layouts. Avoid Word documents with headers, footers, or text boxes — these elements are often skipped or extracted out of order. Avoid mixing topics in one answer — keep each answer focused on one question. Avoid outdated information — the AI will confidently state whatever is in the knowledge base, including old prices, discontinued services, or changed policies.

Question 7

Can I copy a ready-made template?

Accepted Answer

Yes. Copy the template into a plain text file, fill in your own content, and upload it. Save the file as knowledge-base.txt. The template structure is: Question 1: What are your opening hours? Answer: [your opening hours including every day of the week and holiday exceptions]. Question 2: Where are you located? Answer: [full address with directions and parking]. Question 3: How do I book an appointment? Answer: [describe the booking process step by step]. Question 4: What is your cancellation policy? Answer: [cancellation terms, notice required, and fees]. Question 5: What payment methods do you accept? Answer: [list accepted payment methods]. Continue with as many Question/Answer pairs as you need. There is no maximum number of questions in a single file.

Question 8

What about non-Q&A content like service descriptions or policies?

Accepted Answer

For content that is not naturally Q&A format (such as a full service menu or returns policy), you have two options. Option 1 (recommended): Convert it to Q&A by rewriting the content as questions a caller might ask — this gives the AI the best chance of retrieving the right information for each specific question. Option 2: Use paragraph format with clear headings, keeping each paragraph on one topic with all relevant details in the same paragraph (the system splits by paragraph). Q&A format will always outperform paragraph format for call-handling accuracy because each retrieval unit maps directly to a question a caller might ask.

Best File Format for Uploads