Corrections

Corrections is one of the most powerful features available in Attlas PRO. While your knowledge base helps train your chat, corrections allows you to improve specific responses based on real user interactions. Instead of updating documents, files, or content sources every time an issue is identified, you can directly correct individual responses and teach Attlas how to answer similar questions in the future. This makes your chat smarter, more accurate, and continuously improving over time.

---

​

Why are corrections important?

Every correction improves your chat’s ability to serve users. Benefits include:

• More accurate responses;
• Higher Coverage Rate;
• Better user experience;
• Faster improvements without rebuilding your knowledge base;
• Continuous learning from real conversations.

Corrections allow your chat to evolve based on actual user needs rather than assumptions.

---

​

How corrections works

Whenever a response receives a coverage rate below 80%, it may appear in the Corrections tab for review and improvement. 🟠 Responses with scores between 50% and 79% appear in the Orange tab. These responses are generally acceptable but may benefit from additional improvements or refinements. 🔴 Responses with scores below 50% appear in the Red tab. These responses usually indicate missing information, incomplete answers, or situations where the user’s question was not properly addressed. These responses are automatically grouped so you can review and improve them. The goal is simple: identify gaps, provide a better answer, and ensure future users receive more accurate information.

---

​

What information is available?

When opening a correction, you’ll have access to all the information necessary to understand what happened.

​

Conversation context

Before making any changes, Attlas displays the previous messages exchanged in the conversation. This context helps you understand what the user was discussing before asking the question that received a low score. Understanding the conversation is important because the same question may have different meanings depending on the context.

---

​

User question

You’ll see the exact question that received the low coverage score. This allows you to understand precisely what the user wanted to know.

​

Evaluator feedback

Each correction includes feedback from an external evaluator. This feedback explains why the response received a low score and what information was missing. Examples:

• The response was incomplete;
• Important details were omitted;
• The information was outdated;
• The answer did not directly address the question.

This guidance helps you identify exactly what needs to be improved.

​

Response editor

After reviewing the context, question, and evaluator feedback, you can write an improved answer directly in the correction panel. This new answer becomes the preferred response for similar questions in the future.

---

​

What happens after saving a correction?

When a correction is saved, Attlas creates a new layer of knowledge specifically for that question. From that moment on, whenever a user asks a similar question, Attlas will prioritize the corrected answer. This means the corrected information takes precedence over outdated or incomplete information that may still exist in your knowledge base. You don’t need to update every document or retrain your entire chat to fix a specific issue. The correction becomes the new source of truth for that interaction.

---

​

Best practices

To get the best results from corrections:

• Review low-coverage responses regularly;
• Read the evaluator feedback carefully;
• Consider the conversation context before editing;
• Provide complete and up-to-date answers;
• Update your knowledge base when recurring issues appear.

---

​

Turn mistakes into improvements

No chat is perfect from day one. Corrections helps you transform weak responses into learning opportunities, ensuring your chat becomes more accurate with every interaction. The more you review and improve your responses, the better Attlas becomes at helping your users.