When a visitor types a question into the Ask AI tab, Helpdesky:
- Embeds the question into a vector
- Finds the most relevant chunks from your published articles
- Sends those chunks + the question to your chosen LLM
- Streams a grounded answer back, with links to the articles it used
This is the standard retrieval-augmented generation (RAG) pattern — your visitors never get hallucinated answers from outside your docs.
Supported providers
- Anthropic Claude — recommended for conversational tone
- OpenAI GPT-4o / GPT-4o-mini — fastest streaming
- Google Gemini — best for long-context article sets
- xAI Grok, DeepSeek, Qwen, Mistral, Perplexity — bring your own key
Open Dashboard → AI → Providers, paste a key, and pick a default model. Keys are encrypted at rest.
Privacy: Helpdesky never trains on your articles. We send each question to the provider you chose, with only the chunks needed to answer it.
Tuning the answers
- Pin your 3–5 most important articles so they always score high
- Mark internal-only articles hidden from widget
- Add a short excerpt to each article — it's a strong signal to the embedder
- Re-index a single article after editing from the article toolbar