AI answer quality depends on two things: whether retrieval finds the right paragraphs, and how you frame the question. The Playground lets you compare different question styles, models, and retrieval ranges — no code required.
Key takeaways
By default the AI is instructed to ground answers in your KB, cite sources, and admit when it doesn't know — the safest setup, avoiding hallucinations. Want answers softer (allow general knowledge) or stricter (literal citations only)? Switch styles in the right-side settings panel — no prompt engineering needed.
After each query, the right-side Sources panel lists what the AI actually drew from: text, filename, similarity score. When an answer is wrong, check these first — most failures are retrieval misses (rephrase and you'd find it), not the AI making things up.
Toggle the model selector at the top of the Playground (Claude Haiku for speed, Claude Sonnet for reasoning, GPT-4o for general accuracy, Gemini Flash for multimodal). Run the same question and compare. Rule of thumb: Haiku for quick Q&A, Sonnet for complex analysis, Gemini when images or tables matter.
Retrieval Count defaults to 5 — concise answers but may miss details; bump to 8-10 for thoroughness (slower and costlier). Temperature defaults to medium: for factual questions (numbers, rules, dates) set it to Precise to reduce errors; for AI suggestions or copywriting, set it to Creative for richer output.