2. Connecting Data

Klariton answers are only as good as the sources behind them. Connect the strongest sources first, then expand.

Source Types

Klariton uses two broad source types:

• Product data: structured catalog facts such as name, description, vendor, product type, price, availability, images, tags and custom fields.
• Material: unstructured knowledge such as PDFs, DOCX files, Markdown, text files and crawled URLs.

Together they form the knowledge base used for BIQ generation, chat answers, product recommendations and quality checks.

Which Source First?

Situation	Start with
Shopify shop with maintained products	Shopify connection, then one or two official guides/FAQs
Non-Shopify catalog	CSV/API product import, then material upload
Weak product data but strong editorial content	Material first, product data second
High-risk claims	Official approved material only

Shopify LIVE

Shopify connects through OAuth. Klariton requests read-only access for product and catalog information. Klariton does not write products, orders or customer records back to Shopify.

Imported product information includes product names, descriptions, vendor, product type, categories, tags, images, prices, currency, availability and metafields.

Keep the connection active. Re-sync marks changed products as stale so you know when answers may need regeneration.

CSV/API Products LIVE

For CSV/API imports, maintain at least:

• Stable product ID or SKU.
• Name/title.
• Description.
• Category/product type.
• Properties or metafields.
• URL/handle.
• Locale, if known.

Price and availability are useful only if kept current. If they are stale, avoid publishing firm price or stock commitments.

Materials LIVE

Supported upload formats:

• PDF
• DOCX
• Markdown
• TXT

URL crawling extracts the readable main content and stores a snapshot. Crawls may fail if a page requires login, uses bot protection, blocks by geography or exposes only client-side content.

Material Processing

Each material goes through:

1. Parsing and extraction.
2. Chunking into useful sections.
3. Embedding for semantic retrieval.
4. Inclusion in the knowledge base.

Possible states include queued, processing, processed, partially processed and failed.

Quality Checklist

Before importing:

• Remove outdated versions.
• Resolve contradictory product claims.
• Prefer official sources.
• Avoid customer personal data, orders, payments and raw support dumps.
• Keep languages separated.
• Use clear headings in long documents.

If Processing Fails

Try:

1. Use a supported file type.
2. Export scanned PDFs to real text.
3. Remove huge tables and irrelevant appendices.
4. Test crawled URLs in a clean browser session.
5. Upload a cleaned document if a page blocks crawlers.
6. Regenerate affected BIQs after adding better sources.