What "training" actually means (and doesn't)
No mainstream website chatbot fine-tunes an LLM on your content — fine-tuning is slow, expensive, and frozen the moment your pricing page changes. What products call training is indexing: your pages are split into chunks, embedded as vectors, and stored so the relevant pieces can be retrieved and handed to the model per question. That's retrieval-augmented generation — the full mechanics are in what is a RAG chatbot. The practical consequence: the quality of a "trained" chatbot is the quality of its ingestion pipeline, not of some bespoke model.
The three ingestion paths, ranked by ongoing effort
- Manual uploads (PDFs, docs, pasted text): full control, permanent chore. Right for content that isn't on your site — internal policies, spec sheets. Flocci Chat supports these from the dashboard.
- Scheduled crawling: point the bot at your sitemap, re-sync daily/weekly. Fine for static server-rendered sites; fails on JavaScript-rendered content and lags every content change until the next sync. Many tools also cap indexed pages per pricing tier.
- Crawl + passive learning (automatic, Flocci Chat's default): an onboarding crawl covers day one; from then on the embedded widget reads every page a real visitor opens — after scripts run — and updates the index. Changed a page? It re-learns on the next visit. New page getting traffic? Already indexed. No caps, no schedule, no chore.
How ingestion quality shows up in answers
Behind the scenes, Flocci Chat chunks content section-aware (headings, code blocks, FAQs get appropriate profiles), deduplicates near-identical chunks, and retrieves with hybrid vector + full-text search reranked for diversity — so an answer about "refund timelines" pulls your actual policy paragraph, not four copies of your footer. You can inspect exactly how any document was chunked and test retrieval in the dashboard's diagnostic tools before going live.
Keeping it fresh — the part everyone underestimates
Support content changes weekly: prices, features, policies. With scheduled-crawl tools, every change opens a staleness window where the bot confidently recites your old pricing. With passive learning the window closes on the next page visit — and your own five-minute browse after publishing closes it immediately. Pair that with the dashboard's unanswered-intents report (what visitors asked that your content couldn't answer) and "training" becomes a feedback loop: publish → auto-learn → see gaps → publish.