“Hey ChatGPT - can you summarize this article for me?” Queries like this dominate the AI sphere as academics and professionals alike rely on AI to explain long-form content.

This prompt may seem straight forward at first, but in reality there are aspects of it that are unclear. How in depth should this analysis be? Does the user want an overview, or a technical analysis of the article?

These aren’t questions that can be answered unless specified by the user. LLMs are trained to respond plausibly when given clear direction - but many user prompts are unintentionally and subtly ambiguous. The same under-specified user prompt may yield an analytical report or a list of bullet points.

This project explores a design choice to complement AI systems: a pre-query filtering system to collect context from large text documents. By aggregating and injecting context from the document into the user’s prompt, the LLM is implicitely biased towards responses that more closely reference the source material. All of this can be done without further input or fine tuning from the user.