NikTech AI

Why these kits exist. Before building, I cataloged 21 RAG explainers, courses and videos and scored each with a decision model on four questions: does it teach evaluation mechanics (golden set, retrieval hit vs citation hit, the judge, the gate), does it teach retrieval mechanics with numbers (threshold, BM25 + cosine fusion, top-k), is it usable from zero, are the numbers from a real run.

No external resource scored above 0.5 on both mechanics columns. The two that teach evaluation are code courses for people who already know the words. The free RAG visualizers all stop where retrieval ends.

The method, the full table (theirs and mine, same questions) and the raw files are public: niktechai.com/blog/rag-explainers-scored

If you find something I missed, tell me and I will add it to the table.

One answer from the kit, free, since it is the question every panel asked me this month: how do you pick the similarity threshold in RAG?

1. Guess from the embedding model's score bands. Unrelated text lands near 0.5, related near 0.7.

2. 2. Turn the floor off. Run five real questions and five junk ones.

3. 3. Read the lowest real score and the highest junk score.

4. 4. Set the floor just above the junk. Write the reason next to the constant.

5. 5. Re-measure when the model or the notes change.

On my own store the guess was 0.66; the measured piles (real 0.74 to 0.86, junk 0.67 to 0.72) made it 0.72.

That is question 10 of 46 in the GenAI Interview Kit. Each one is written the same way: symptom, detect, cause, fix, prevent, with the mechanism drawn out.

Welcome. Two study kits are live here, both built while I did the same interviews and the same exam in September 2026.

GenAI Interview Kit ($39): five interactive HTML pages. 46 interviewer questions with full model answers, 45 troubleshooting cards in symptom, detect, cause, fix, prevent order, 42 code blocks, a RAG program and an agent program walked through in 30 frames. Works offline, no login.

AI-103 Flight School ($49): 68 slides, 94 interactive widgets (steppers, hotspots, sliders, guess-before-reveal drills), all five domains, a 40-question simulator with an explanation for every answer. Every question original, written from the published skills outline.

Bundle: both for $69.

Try first: the free product on the Products tab has a full kit page and the AI-103 exam blueprint, no card needed. Or open them straight away: niktechai.com/kits

Questions here or hello@niktechai.com.