Source ingestion

Supported file types, URL ingest, and the safety checks every source goes through.

Ingest is how raw material — text, a URL, or a file — becomes wiki pages. A single ingest run typically produces 10–15 new pages and strengthens related existing pages.

Supported input types

Text & Markdown

Paste text directly, or upload .txt / .md files. Streamed over SSE as the wiki is built.

PDF

Text-layer extraction by default. If a scanned PDF has no usable text layer, it can fall back to OCR when enabled.

Word & Excel

.docx and .xlsx are extracted directly — no manual conversion needed.

URL

Fetch and ingest a web page. Every URL is validated against SSRF (server-side request forgery) protections before it’s fetched.

Images

Extracted with Claude vision by default, or with zero-cost local OCR when your tenant has OCR enabled.

Audio

Transcribed in-browser (Web Speech API) on the Vessios frontend, then ingested as text — no audio file upload required.

How ingest works

Before any content reaches your wiki, every ingest goes through the same pipeline:

SSRF validation — for URL ingest, the target is checked before it’s ever fetched.
Duplicate detection — content is hashed (SHA-256) to avoid re-ingesting the same material.
Prompt-injection screening — ingested content is scanned for attempts to manipulate downstream LLM behavior.
PII flagging — content flagged as containing personal data is isolated from search until an admin reviews it (see Security).
Extraction & page generation — the source is turned into 10–15 new or strengthened wiki pages, plus wikilinks connecting them to related pages.

Large uploads (files, and multi-step sources) go through an asynchronous two-stage flow: the file is uploaded to a signed storage URL, then a background job performs extraction and ingest. Small text and URL ingests stream progress back over Server-Sent Events (SSE) so you can watch the wiki build in real time.

Confidence and review

Extraction isn’t always perfectly confident — a low-confidence scan, for example, may need a human to confirm it. Vessios automatically routes low-confidence extractions to an Async Review queue rather than writing them into the wiki unreviewed.

Quotas

How much you can ingest depends on your plan — see Plans & billing for exact limits on ingests per month and per day.