Custom GPT doesn't read the knowledge

[fervent mulch]

I’ve created a custom GPT and uploaded a .json file as its knowledge source. However, when I ask questions, it tends to respond with information from the first topic it encounters in the documentation—even when the correct answer appears later, deeper in the file.

How can I ensure the custom GPT searches through the entire documentation before providing a response?

I’ve already tried writing specific instructions to encourage deeper searching, but so far, that hasn’t been successful. For example:

• When a term or concept is introduced in multiple parts of the documentation (e.g., module overview and detailed feature sections), always prioritize the most specific and practical explanation, not the first occurrence.
• Trigger: User requests a definition
• Instruction: Search the documentation thoroughly and prioritize the sections where this feature is most frequently mentioned or described in detail. Take additional time to analyze areas with high concentration of related terms. Once located, provide a clear, step-by-step explanation based strictly on that content.

[south radish]

hi

[fervent mulch]

hi

[twin heath]

I'm running into the same issue, but it seems like it's reading partially. My use case is I'm trying to return s3 path for additional knowledge via API call, and it's getting the directory wrong by a single letter.

No amount of prompt manipulation has managed to get it to stop doing this. I'm thinking I'm going to shift the entire thing to search capability via API instead and see if I get better results that way