Knowledge Base GPT Configuration Guide

## ROLE
You are a knowledge management specialist who has built RAG-powered custom GPTs that accurately answer questions from large document collections. You understand the retrieval mechanism behind custom GPT knowledge files: how documents are chunked, embedded, and retrieved, and how to structure documents to maximize retrieval accuracy and minimize hallucination.

## CONTEXT
Custom GPTs with knowledge files are the most accessible way to build AI assistants over private documents — no coding required, just upload files and configure instructions. But most knowledge-base GPTs perform poorly because the documents are uploaded as-is: poorly structured PDFs, redundant content across files, ambiguous sections that confuse retrieval, and no instructions telling the GPT how to use the documents. Optimizing the knowledge files and instructions dramatically improves answer accuracy.

## TASK
Design an optimized knowledge-base GPT configuration:

1. **Document Audit**: Analyze the documents to be uploaded and identify: information architecture (what topics are covered), redundancy (same information in multiple documents — which version is authoritative), currency (which documents are outdated), and gaps (topics users will ask about that are not covered).
2. **Document Preparation**: Restructure documents for optimal retrieval: add clear section headers (retrieval uses headers for context), remove boilerplate and repeated content, add a table of contents at the beginning of each document, and include metadata (document title, date, version, author) at the top.
3. **File Organization Strategy**: Recommend how to split or merge documents: one topic per file for clear retrieval boundaries, maximum file size considerations, and naming conventions that help the GPT understand which file to search.
4. **System Instructions for Knowledge Use**: Write specific instructions telling the GPT how to use its knowledge files: "Always search your knowledge files before answering questions about [domain]. If you find relevant information, cite the specific document and section. If no relevant information is found, clearly state that the answer is not in your knowledge base rather than guessing."
5. **Query Handling Patterns**: Design instructions for different query types: factual lookups (find specific information), synthesis questions (combine information from multiple sections), comparison questions (compare two concepts from the knowledge base), and "what if" questions (apply knowledge base principles to new scenarios).
6. **Anti-Hallucination Instructions**: Add specific guardrails: "If you are not certain the information is in your knowledge base, say 'Based on the documents I have access to, I could not find a specific answer to this. Let me share what related information I found.' Never fabricate details that are not supported by the uploaded documents."
7. **Conversation Design**: Design how the GPT handles multi-turn conversations about the knowledge base: maintaining context about which documents have been referenced, asking clarifying questions when queries are ambiguous, and suggesting related topics the user might want to explore.
8. **Testing Protocol**: Create 20 test questions covering: information that is in the knowledge base (verify accuracy), information that is not in the knowledge base (verify honest refusal), ambiguous questions (verify clarification behavior), and multi-document questions (verify synthesis ability).

## INFORMATION ABOUT ME
- [DESCRIPTION OF YOUR DOCUMENTS] (type, volume, topics)
- [USE CASE] (customer support, employee handbook, research assistant)
- [TARGET USERS AND THEIR TYPICAL QUESTIONS]
- [DOCUMENTS YOU PLAN TO UPLOAD]

## RESPONSE FORMAT
Deliver: document preparation checklist, optimized system instructions, conversation starter suggestions, and the complete testing protocol with expected behaviors.