Witivio Documentation

# Optimized Instructions Writing for GPT Pro

A high-performing chatbot relies on clear, precise, and well-structured instructions.
If instructions are vague or incomplete, responses will be inconsistent, inaccurate, or off-topic. Add system Prompt in GPT Pro: To add a system prompt, use 'System Prompt' tab. Write a prompt and think to click on 'Save'.

Prompts

This guide helps you write optimal instructions for your GPT-4o chatbot, ensuring accurate responses and strong alignment with business expectations.

# 1️. Structuring Instructions Effectively

# 1.1. Structure:

# Why is structure important?

GPT Pro follows instructions based on logical structure. A clear format helps prevent ambiguity and ensures consistent, reliable behavior.

## Objective  
Define the chatbot’s role and covered topics.
## Tasks  
Clearly list what the chatbot can and cannot do.
## Information Sources  
Specify where the AI should search for answers.
## Process  
Describe the steps the AI must follow to handle a request.
## Behavior Rules  
Set tone, rephrasing guidelines, and error handling.
## Example Conversations  
Include sample dialogues to guide the AI.

# 1.2. Example:

## **Purpose**
You are **Lexi RH – Payroll & Time Management Assistant**, an AI specialized in payroll and time & attendance topics.  
You assist employees of **Company X** with HR-related queries.

## **Tasks**  
- Answer questions related to **payroll, leave, time tracking (GTA), and HR processes**  
- Guide employees through **internal tools** (leave requests, time entry, approval workflows, etc.)  
- **Maintain conversational context** when users follow up on previous questions  
- **Smartly rephrase vague queries** to improve search results  
- **Always ask for clarification** in case of ambiguity or multiple possible answers  

## **Knowledge Sources**  
- **HR Snippets** (validated Q&A pairs by the company)  
- **Witivio documentation base** (Payroll & Time Management)  
- ⚠️ **Do not use generic LLM knowledge** — only rely on the provided sources

## **Capabilities**  
- Search through validated HR documents  
- Understand and track **conversation history** and follow-up questions  
- **Proactive clarification** behavior:  
 - Ask for clarification when vague terms are used (e.g., "bonus", "absence", "comp time")  
 - If multiple interpretations are possible → **Always ask the user to specify their intent**

## **Workflow**  
1. Check **HR snippets** first  
2. If no match, **search the documentation base**  
3. If the user follows up on a previous message → **rephrase the full query and search again**  
4. If the term is vague or multiple results are found → **ask the user to clarify**  
5. Generate a **clear, concise, and context-aware answer**  
6. If no answer is found, display:  
  > "I'm sorry, I don't have an answer to your question.  
  > If you'd like to open a ticket, click 👎 in the reference card and then select *Open a ticket*.  
  > You can also leave feedback to help improve the service."


## **Example Conversations**
### Case 1: Contextual Follow-Up in Time Management

**User**: "How do I enter a work exemption in GTA?"  
**Lexi RH**: "Open [tool name], go to 'Time Management' > 'New Entry', select 'Work Exemption', and validate."  
**User**: "What's the function name?"  
**Automatic Rephrasing**: *"What is the function to enter a work exemption in GTA?"*  
**Lexi RH**: "The function to use is [function name]."

### Case 2: Vague Request  
**User**: "What are the rules for a bonus?"  
**Lexi RH**: "Could you specify which type of bonus you're referring to? (seniority bonus, performance bonus, exceptional bonus…)"

### Case 3: Multiple Possible Matches  
**User**: "How do I request a day off?"  
Several absence types detected (paid leave, RTT, unpaid leave...)  
**Lexi RH**: "There are different types of absences. Are you referring to paid leave, unpaid leave, sick leave, or something else?"  
Once clarified, Lexi RH provides the relevant answer.

## 2️. Writing Clear and Precise Instructions (Max 4000 characters)
### Why precision matters:  
- Define the business context: Who uses the AI? What questions are expected?  
- Specify allowed and forbidden actions to avoid irrelevant answers  
- Explain the response process: Where and how the AI should search  
- Frame tone and rephrasing logic: When to clarify or ask for more details
💡 **Tip:** If the AI must display a specific error message, **put it in quotes** to prevent rephrasing or omission.
**Example:**  
If no answer is found, the AI must display:  
*"I'm sorry, I don't have an answer to your question."*

# 3️. Handling Follow-Ups and Rephrasings

# Why is this important?

If users re-ask a question, it usually means the previous answer was incomplete or unclear.

## Process  
1. Check if the answer exists in the FAQ.  
2. If the question is a rewording, search directly in documents.  
3. Rephrase the query to refine the search.  
4. If no relevant result is found, show the predefined error message.

# Why this works:

  • Gives the chatbot a second chance to succeed
  • Improves precision by letting the AI rephrase before failing

# 4️. Test & Iterate to Improve Chatbot Performance

# Why test regularly?

A well-designed prompt should evolve based on feedback.

# Best practices:

  • Run A/B tests with different prompt versions
  • Update instructions based on new business needs
  • ⚠️ Always click “New conversation” after updating instructions to apply changes

# Summary of Best Practices

Objective Action
Structure the instructions Use clear, modular sections
Maximize precision Write detailed and specific instructions
Handle follow-ups smartly Optimize the workflow to reduce failures
Specialize the AI Add domain-specific acronyms and definitions
Test and iterate Continuously analyze and adjust based on feedback

# GPT Pro Agent Optimization & Configuration

# 1. Prompt System Instructions

# Purpose

The prompt system defines how the chatbot works. It controls its behavior, responses, and scope.

# Best practices:

  • Clearly define the chatbot’s role and domain of expertise
  • Control information sources: Only use knowledge base and document corpus
  • Manage rephrasing behavior during follow-ups
  • Request clarification in case of ambiguous terms
  • Define a clear fallback error message:
    "I'm sorry, I don't have an answer to your question. Please contact support at support@contact.com."

# 2. Customizing the AI with Acronyms & Business Terms

# Why this matters

The AI works via semantic matching. Without a business lexicon, it might misinterpret terms.

# Best practices:

  • Add all key business terms and acronyms under "Prompt > Business Vocabulary"
  • Update the list regularly Example:
    In an HR context:
  • DPAE = Declaration of Employment
  • GTA = Time & Activity Management
  • PSS = Social Security Ceiling ✅ GPT Pro format:
    RTT(Réduction du Temps de Travail, rtt)

# 3. Writing Effective Knowledge Base Questions

# Prerequisites:

  1. Configure and publish the KB
  2. Add documents via “Local documents” or SharePoint library

# Why wording matters

The chatbot uses strict matching on question titles (default threshold: 0.12). Only closely worded queries are matched.

# What is a snippet?

A snippet = One KB entry composed of:

  • A well-formulated question
  • An associated answer

# Best practices

  • Use natural, specific question phrasing
  • Avoid too broad or too niche wording Example:
    ❌ "What is the leave policy?" → Too broad
    ✅ "What are the criteria for unpaid leave?" → Specific and precise

# 4. Knowledge Base vs Document Priority

# Source hierarchy:

  • KB is prioritized → If a match exists, documents are ignored
  • Documents are only used if KB fails
  • Exception: If “Use KB as AI Gen context” is enabled → both KB and docs are searched

# 5. Document Vectorization and Chunking

# How document search works

Documents are split into small text chunks and vectorized. When a question is asked, GPT performs a similarity search.

# Best practices

  • Structure documents with clear titles and subtitles
  • Keep documents up to date
  • Use short, well-segmented paragraphs
  • Prefer structured sections over long tables Example:
    ❌ Long table with no headers
    ✅ Short, titled sections with bullet points

# 6. Matching Sensitivity Settings (Default: 0.12)

# 🔁 What is matching flexibility?

Value Effect
0 Strict match only
0.12 Precise match (default)
1 Wide match, less precise answers

# Best practices

  • Keep 0.12 for precise answers
  • Adjust to ~0.15 if some close queries aren’t matched
  • Reduce below 0.12 if AI gives too many approximate answers ⚠️ Matching sensitivity must be adjusted by your Witivio CSM.

# Troubleshooting Guide

# My chatbot isn’t giving the right answer, what should I do?

# Step 1: Identify the source of the issue

  • Wrong answer? → Poor matching in KB
  • No answer? → Missing data in KB or documents
  • Wrong source? → Misconfigured priorities or unclear content Use the “References” button to verify the source.
    If no reference is shown → The AI used its generic knowledge (not allowed)

# Step 2: Check KB Question Wording

  • Test multiple phrasings
  • Use duplication to add variants
  • Add synonyms where necessary
    ⚠️ Click "New conversation" after any change in instructions or vocabulary

# Step 3: Adjust KB & Document Content

  • If document info is better than KB → Unpublish outdated KB entries
  • Add a summarized snippet based on document content

# Step 4: Improve Document Structure

  • Segment content clearly
  • Avoid dense or complex tables

# Step 5: Tune Matching Sensitivity

  • Increase to 0.15 if close matches fail
  • Reduce to 0.10 if too many approximate answers appear

# ✅ Troubleshooting Summary

Step Check Solution
1. Source KB / docs / generic? Adjust priorities or fix source
2. Question phrasing Wording close enough? Rephrase or duplicate the KB question
3. KB vs Docs Better answer in docs? Unpublish KB entry and rely on documents
4. Document structure Info easy to find? Better segmentation
5. Matching sensitivity Too strict or too loose? Tune between 0.10 and 0.15
6. Business vocabulary Are acronyms defined? Add them in prompt → "New conversation"

← FAQ →