Self-checking research assistant

Gemini 2.5 Flash Lite writes the brief from retrieved Wikipedia articles. A second model, Gemini 2.5 Flash at temperature 0, then checks the draft for claims the sources do not support and for the word limit.

Before any retrieval, a separate check rejects inputs that are not industries, so a query like "red" or a string of random letters does not trigger a search.

I tested the app with deliberately awkward inputs as well as the automated check. That led to changes in the retrieval settings and to a system prompt that tells the model to say when information is not available, rather than filling the gap with numbers.

In the write-up I set out what a production version would need: retrieval over a company's own documents from a vector database, locally hosted models where data privacy matters, and caching for repeated queries.