When a bot based on Gen AI (Generative Artificial Intelligence) does not respond correctly or provides incomplete answers (for example, without including the technical sheet or the corresponding URL), one of the most frequent causes is inadequate configuration of training files or a prompt that is not specific enough.
📋 Checklist before training
Before uploading documents to enrich your bot's knowledge base, make sure you meet these points:
✅ Upload only relevant files: Avoid uploading product technical sheets that are not mapped or that are not part of the answers you expect to provide.
✅ Verify the models: Check that all current models or products are included in the file.
✅ Unify the information: It is better to consolidate technical sheets in a single document, rather than dispersing them across multiple PDFs or separate files.
✅ Exact nomenclature: Verify that the product/model name in the document matches exactly what is used in the bot's mapping or in the user's query.
✅ Clean the base: Avoid uploading contradictory or outdated training that could confuse the AI.
💬 How to structure the prompt to deliver technical sheets
If you need the bot to deliver specific documents, you must include very clear examples and restrictions in the prompt to guide its response.
For example, in the [Tasks] section you can include an instruction like this:
👉 "When the user asks about a product, search for the information in the database. If the product exists, you must mandatorily include the main features, the link to the technical sheet in PDF and the official purchase URL."
🚫 Risks of overtraining
Loading too many documents, especially if they are not properly structured or if they contain products that are not part of the actual conversational flow, can cause serious problems:
Generate data hallucinations (the AI invents features that do not exist).
Make the bot prioritize irrelevant information before the direct answer.
Cause empty, confusing or incomplete responses.
✅ Additional recommendations for your Prompt
To further refine your bot's behavior, apply these golden rules in your instructions:
1. Be clear and specific in the instruction: Explicitly indicate what information source the model should consult (e.g., your official database, FAQs or validated documents).
Example: "Consult only the official database to answer this question."
2. Define behavior in the absence of data: Specify what the model should do when it doesn't have the answer: respond with an error message, forward to a human, or admit that it doesn't know.
Example: "If there is no confirmed information in your documents, respond: 'I don't have that information available at this moment, would you like me to connect you with an agent?'"
3. Limit generation to verifiable answers: Explicitly ask the AI to avoid generating speculative responses.
Example: "Don't make assumptions, don't invent data or offer products that are not in the catalog."
4. Use examples and counterexamples: Incorporate examples of correct responses (what you want it to say) and examples of how to respond if it doesn't have the information. This helps the model understand the limits and expected format.
5. Reinforce the importance of accuracy: Add closing phrases in the [Rules] section that remind the AI that its highest priority is to provide accurate and 100% reliable answers based on its training.
Take control of your AI! 🚀
By correctly organizing your documents and fine-tuning your prompts, you ensure that your bot provides accurate and reliable answers. Avoid hallucinations and improve your customers' experience with precise information. ✅