Skip to main content

Create and manage a knowledge base

Start with a small, coherent collection. A knowledge base should have a clear owner, subject, source policy, and update cadence.

Create the collection​

  1. Open Knowledge.
  2. Select + or Add Knowledge Base.
  3. Enter a clear Knowledge Base Name.
  4. Add a Description so workspace members understand what belongs in it.
  5. Add Instructions that tell attached agents how to use the collection.
  6. Select Create.

Description is for identification. Instructions become prompt context when the base is selected in an agent configuration. Keep instructions about source scope, conflict handling, missing answers, and escalation; put the factual material in sources.

Example instructions:

Use this collection only for approved product, delivery, and return information. Prefer the most specific current policy. If the retrieved material does not answer the question or two sources conflict, do not guess; offer support follow-up.

Add supported sources​

Files​

The Knowledge page accepts PDF, TXT, DOCX, CSV, and XLSX files. Each upload can be up to 10 MB, and the full knowledge base can contain up to 50 MB of indexed source content.

You can select a file or drag one onto the page. Wirevox extracts text, organizes it, creates chunks, embeds them, and displays the result under Indexed Sources.

Image-only PDFs can fail because the Knowledge page PDF path requires extractable text. Convert them with OCR first or upload the page image through Vox so its approved visual analysis can be stored as a text source.

Public website pages​

Single Page imports one exact public HTTP(S) page. Entire Website scans the starting page for internal links, with a sitemap fallback when few links are found, and lets you select up to 20 pages for import. Scanning can discover up to 100 candidate links; it does not automatically index every page.

Only selected pages are crawled. Wirevox removes common navigation and layout elements and extracts readable page text. Heavily JavaScript-dependent, authenticated, blocked, or low-content pages may fail.

Private-network URLs, localhost addresses, embedded credentials, and non-HTTP(S) URLs are not allowed. Redirects are checked and limited.

Text documents​

Select Create document to write or paste normal text. The current Knowledge page requires at least 200 characters and recommends placing shorter facts directly in an agent prompt or variable.

Use descriptive source names. A repeated source name replaces the existing chunks for that source during re-ingestion.

Vox attachments and research​

Vox can prepare an approved knowledge action using uploaded supported files, analyzed images, exact public URLs, or normal text documents. Its composer currently accepts up to two staged attachments at once; the files are not added merely by uploading them.

For an image, Vox stores the approved analysis text rather than the bitmap itself as searchable knowledge. Review important names, numbers, tables, and small print before approval.

Vox can also research public pages without modifying knowledge. To persist that research, explicitly ask it to create a knowledge base or add sources, review the proposed source URLs and text, and approve the action.

Monitor processing​

URL imports begin in the background and appear as Processing until they succeed or fail. You can leave the page while processing continues. Do not attach or publish a critical workflow until the required sources show successful chunk counts.

The source table shows source name, type, indexed size, chunk count, and last-crawled time. Search filters this table; it does not test semantic retrieval.

Inspect and edit chunks​

Open a source's chunks to inspect what the model can retrieve. You can edit a chunk or delete it. Saving an edit creates a new embedding for that content. Deleting the final chunk removes that searchable content from the source.

Inspect chunks when:

  • a document contains complex tables or columns;
  • a website repeats navigation or legal boilerplate;
  • a PDF uses unusual layout;
  • a source is very long;
  • exact prices, dates, model numbers, or contact details matter; or
  • retrieval tests return incomplete or unrelated answers.

Update or remove sources​

Use Update for selected URL sources to crawl them again. Re-crawling replaces chunks associated with the exact URL. Uploaded files and text sources should be re-added with the same source name when replacement is intended.

Deleting a source removes all of its chunks. Wirevox also removes that source name from reusable Knowledge functions that explicitly selected it. Review those functions afterward: an empty source selection means the function searches the entire knowledge base.

Deleting a knowledge base is broader and permanent. Confirm attached agents and Knowledge functions before deletion.

Source-quality checklist​

  • Use authoritative, current material.
  • Separate incompatible brands, jurisdictions, or policy versions.
  • Remove obsolete duplicates.
  • Make each document understandable without hidden visual context.
  • Include the terms customers actually use.
  • Avoid secrets or information the agent is not permitted to reveal.
  • Record an owner and review date outside the source content or in a clearly maintained policy document.

Next step​

Continue with Attach knowledge and choose retrieval scope.