Skip to main content

​
Knowledge

Knowledge is trusted source material your app can search. Keep the real records (chapters, policies, articles) as normal tables. GenieForge owns the search copy: chunking, indexing, citations, and a plain-text fallback while embeddings are still catching up. Ask Build to attach knowledge to a table you already have. Example: search chapters by their body text, then open the matching chapter. You can also attach knowledge to documents that store a file (a PDF or a text upload). Search can filter by fields on that document, such as category.

​
What you get

  • A named collection you can search from chat or a page.
  • Results that point back to the original row, not a mystery snippet, including fields like title and category when they exist.
  • Search that uses plain text over the full collection while embeddings are still catching up, so later sources are not missing from results. That fallback matches the words in your query, not only the exact phrase.
Indexing that search copy uses credits, the same way other AI usage does. Looking up a search query does not add a separate charge. If the app is out of credits, new text waits to be indexed and search stays on the plain-text fallback until credits are available. You do not create a separate chunks table or ask the AI to invent a retrieval pipeline.

​
After source rows change

When you add or edit the source records, ask Build to sync that knowledge collection. Sync replaces the search copy in one pass so a half-finished ingest cannot wipe a known-good set. If a file-backed document is still extracting, the previous search copy for that document stays until extract succeeds.

​
If you already built a chunks table

If an older app created its own chunks table (including one that points at documents with a different id column), ask Build to declare knowledge on that table instead and remove the custom ingest tool. See Data structure.