Training your chatbot — supported and safe sources

How training works

Pivra extracts text, splits it into searchable chunks, and creates embeddings. When a visitor asks a question, the most relevant chunks are supplied to the model as context. Adding a source does not change the chatbot's core security instructions.

Use material that the business owns or is authorised to use. Do not upload passwords, secret keys, private customer records, health information, payment details, or documents you would not want an authorised workspace member to read.

Supported sources

Website URL

Use this for one important public page. Pivra validates public HTTP/HTTPS destinations, applies fetch and content limits, and rejects unsafe network locations and non-text responses.

Sitemap

Use a public sitemap.xml to train multiple pages from the same website. Sitemaps are the best option for Pivra's own Help Centre because new articles can be discovered without uploading another file.

YouTube

Paste a public YouTube URL. Training succeeds only when a usable transcript is available.

Files

The dashboard accepts PDF, DOCX, TXT, Markdown, and CSV files up to 10 MB. Pivra verifies the file signature rather than trusting only the filename. Password-protected, malformed, oversized, unsupported, or suspicious files are rejected.

When the private malware scanner is configured, uploads must pass that scan before text extraction and embedding. A successful scan reduces risk but does not make untrusted document claims true; review the extracted business information before publishing.

Plain text

Use plain text for short, controlled facts such as service areas, approved FAQs, escalation wording, and policies. Include a descriptive title so workspace members know who owns the information.

Source status

Do not publish based on a source until it is Ready.

Limits

The dashboard enforces both the 10 MB upload limit and plan-level source/character limits. Current plan allowances are shown under Billing & Usage and on the public Pricing page. These are the canonical values if an older article or screenshot differs.

Keep knowledge current

Website, sitemap, and YouTube sources can be set to no automatic retraining, daily, weekly, or monthly retraining. Use:

• daily only for genuinely fast-changing public information
• weekly for active service and support sites
• monthly for mostly stable documentation
• manual for policies or pricing that require approval before they change

After a material change, retrain and test representative questions before relying on the updated source.

Improve answer quality

• Prefer a small number of authoritative sources over duplicated or conflicting pages.
• Remove outdated documents instead of leaving contradictory versions active.
• Put critical limitations and escalation rules in business instructions as well as knowledge.
• Use Strict mode when answers must stay within approved material.
• Use Refinements for exact, high-risk wording.
• Review real conversations for gaps; do not train blindly on visitor messages.

If a source fails

Check the failure message first. Then confirm that the URL is public, the file is an accepted type under 10 MB, the document is not password protected, and the workspace has remaining source and character capacity. If a JavaScript-heavy page extracts little text, try its sitemap or paste approved content as plain text.