Knowledge

Your website becomes the starting point.

Engage discovers the pages a customer would look for, extracts what they actually say, and makes those passages available to your agent. You start with what you have already written, not with a blank knowledge base.

Inside the index

Every page, and what happened to it.

This is the shape of what Engage records for a crawl. Each page carries the state the ingest ledger gave it — nothing is summarised into a score.

KnowledgeLast processed 2 minutes ago
Services
  • Emergency repairsIndexed
  • InstallationIndexed
  • Maintenance plansIndexed
Company
  • Service areasIndexed
  • AboutFetchedNo extractable content
  • CareersExcludedOff-topic for the agent
Support
  • FAQIndexed
  • ContactIndexed
  • DownloadsFailedFetch returned 403
  • Terms (PDF)Indexed
  • Price list (XLSX)ExcludedA spreadsheet is not read
  • Press archiveDiscoveredNot fetched: over the page limit
From URL to answer

What processing actually does.

Plain language first. The technical detail sits underneath, for anyone evaluating this properly.

  1. Links are followed within your own domain. A URL policy resolves DNS first and refuses private ranges, so the crawler cannot be aimed at an internal network.

    24 URLs discovered

  2. Pages, and the PDF, .docx and text or Markdown files your site links to, are retrieved up to a configured limit. Failures are recorded against the URL that failed rather than dropped.

    19 pages fetched

  3. The HTML is parsed into a DOM behind a nesting-depth guard, and readable content is separated from navigation, scripts and boilerplate. A PDF, .docx or text file has its text read directly. A scanned PDF with no text layer is not read, because there is no OCR, and a spreadsheet is not read.

    Readable text per page

  4. Long pages are split into passages small enough to retrieve precisely and large enough to keep their meaning.

    Passages with their source URL

  5. Each passage is stored with a vector so it can later be matched to a question by meaning rather than keyword.

    vector(1536) per passage

  6. At question time the closest passages are selected and handed to the agent as the only material it may answer from.

    Top passages + source URLs

Every stage is recorded. When a crawl half succeeds, both halves are reported — a partial result is never shown as a clean one.

Source transparency

Every supported answer names the page it came from.

This is product behaviour, not an illustration. The public API contract requires a supported answer to carry at least one citation and an unsupported one to carry none.

KnowledgeIndexed 2 minutes ago
Services
Company
Support

Do you offer weekend service?

Emergency callouts run on Saturdays until 4pm across Miami-Dade. Standard appointments are weekdays only.

Answered from Services → Availability

When no passage supports the question, the answer comes back with no citations and says so. That pairing — evidence or nothing — is the contract.

Control over the index

What you can shape today, and what you cannot.

In the product
  • Same-origin boundaryAvailableOnly the domain you submit is read
  • Page limitAvailableA large site cannot become an unbounded crawl
  • Recorded exclusionsAvailableEvery skipped URL keeps its reason and stage
  • Ingest ledgerAvailableAttempts, phases and outcomes are append-only
  • Re-indexing on a schedulePartly availableRuns through an operator endpoint, not from the console
  • Approving or removing sourcesPartly availableYou can see each source and remove it; nothing is reviewed before it answers
  • Answers you writeAvailableA question and its answer, written in the console
Not built
  • Reviewing knowledge in the consoleNot built yetThe console area is named but not built
  • Uploading documentsNot built yetThere is no upload path
  • Product catalogue or CRM as a sourceNot built yetNo product catalogue or CRM ingestion

Reviewing and curating knowledge from the console is real work that has not been done. Today the crawl is bounded by policy rather than steered by you.

For technical evaluation

The parts an engineer will ask about.

What exactly is stored?

The extracted text of each page, its canonical URL, a title and a content hash, plus the passages that text was split into and one embedding vector per passage. Source URLs are kept precisely so an answer can cite the page it used.

Which embedding and generation models?

The model boundary is provider-neutral in the domain code, and the validated configuration currently admits one provider, selected by environment variable. Swapping it is a configuration change rather than a rewrite.

How is one customer's knowledge kept from another's?

Every knowledge row is workspace-scoped, and the foreign keys are composite — a row cannot reference a parent in a different workspace, because the database will not accept it. It is a schema property rather than a filter in application code.

What happens to a demo's knowledge when it expires?

It is deleted by a sweep, and the embeddings go with it by cascade. A demo cannot name a durable agent — the schema forbids it — so the sweep structurally cannot reach a customer's knowledge.

Is there a knowledge health score?

No, and we would rather not invent one. The ledger records counts per stage, and a stage that was never reached records nothing at all. Rendering that as a percentage would turn an unknown into a number.

Next step

See what it finds on your site.

The crawl is bounded and reads only what is already public. You will see which pages indexed cleanly and which did not.

Index your website