AI & Knowledge Base
Supported content formats
Every file type and source Liya can index into your knowledge base.
3 min read Updated July 2026
File uploads
| Format | Extension | Max size | Notes |
|---|---|---|---|
| Plain text | .txt | 10 MB | UTF-8 encoding required |
| Markdown | .md, .mdx | 10 MB | Headings, lists, code blocks preserved |
| 10 MB | Text extraction only; scanned PDFs require OCR (see below) | ||
| Microsoft Word | .docx | 10 MB | Formatting stripped; headings preserved as structure |
| Microsoft Excel | .xlsx | 5 MB | Each sheet indexed as a separate section; useful for FAQs in table form |
| CSV | .csv | 5 MB | Best for structured Q&A data; first row treated as header |
| HTML | .html, .htm | 10 MB | Navigation, scripts, and style tags stripped automatically |
| EPUB | .epub | 10 MB | Chapters split into separate indexed sections |
Scanned PDFs: If your PDF contains scanned images of text rather than actual text characters, Liya cannot extract content without OCR. Convert scanned documents to text-based PDFs using a tool like Adobe Acrobat or an online OCR service before uploading.
URL crawling
| Source type | Support level | Notes |
|---|---|---|
| Public HTML pages | Full | Boilerplate, navigation, and ads stripped automatically |
| Public Notion pages | Full | Headings, lists, and callouts preserved |
| GitHub README / wiki | Full | Public repos only |
| Intercom help center | Full | Crawls all public article URLs |
| Zendesk help center | Full | Crawls all public article URLs |
| HubSpot knowledge base | Full | Public articles only |
| Pages behind login | None | Use file upload or native integration instead |
| JavaScript-rendered SPAs | Partial | Server-rendered content only; client-only content may be missed |
| PDFs served via URL | Full | Downloaded and processed as a file upload |
Native integrations
These integrations sync content automatically — no manual uploads needed:
| Integration | What syncs | Sync frequency |
|---|---|---|
| Notion | Selected pages and databases | Real-time via webhook |
| Google Docs | Selected documents and folders | Real-time via webhook |
| Confluence | Selected spaces and pages | Every 30 minutes |
| Intercom Articles | All published articles | Every 24 hours |
| Zendesk Help Center | All published articles | Every 24 hours |
| HubSpot Knowledge Base | All published articles | Every 24 hours |
| Gitbook | Selected spaces | Every 6 hours |
Content that works best
Regardless of format, the content quality matters more than the file type. These characteristics produce the highest AI accuracy:
- Short, focused articles — one topic per article, 100–400 words
- Plain language — write as if explaining to a new customer
- Specific answers — include concrete steps, timelines, and thresholds (e.g. "refunds take 5–7 business days")
- Question-formatted titles — "How do I cancel my subscription?" rather than "Cancellation policy"
- No stale content — outdated pricing or deprecated features confuse the AI; archive old articles
Content size limits
When a single file exceeds the optimal chunk size, Liya splits it into sections for indexing. Splitting happens automatically at heading boundaries (h1, h2, h3) or at 1,000-word intervals when no headings are present.
Very large documents (e.g. a 200-page policy manual) are better split manually into topic-specific articles before uploading. This gives Liya finer retrieval precision.