The problem
A reader with one question — does he know Oracle? what is the CGPA? — has to skim the whole site. A chatbot on a hosted model costs money per visitor and can make things up.
How it works
- 01Split into passages
- 02Search (BM25)
- 03Gemini writes from those only
- 04Answer with source
What was hard
- Short résumé facts and long articles are indexed separately, so one question finds the one passage that answers it.
- A synonym map links words people ask with (“tech”) to words the site uses (“Java 17, TestNG”).
- If nothing matches well, it says so instead of guessing — and never asks the model.
- The Gemini key stays on the server. The browser sends only the question, so the endpoint cannot be used as a free AI proxy; it is rate-limited and cached.
- If the model is off or slow, the best passage is shown as-is. The assistant never depends on the model.
The result
Answers in plain words, grounded in what is on this site, with a link to where each came from.