adaptcxadaptcx

Appearance

Knowledge Bases ​

Knowledge bases give your bot the ability to answer from your own content instead of relying on the LLM's training data. Upload documents (or connect external sources like SharePoint, Google Drive, Confluence), and the platform chunks, embeds, and indexes them for retrieval.

Concepts ​

Knowledge Base (KB) — a named collection of documents your bot can query. Bots reference KBs from RAG Query or KB Search nodes.

Document — an individual file or article in a KB. Documents come from three sources: direct upload, external-source sync, or programmatic ingestion via API.

Chunk — the platform splits each document into ~200–500 token chunks. Chunks (not whole documents) are what retrieval matches against.

Embedding — each chunk is embedded via the tenant's embedding model. Embeddings live alongside chunks in the database's vector column.

Retrieval — at query time, the user's question is embedded and vector-similarity search finds the top K matching chunks.

Getting started ​

  1. Create a KB — Operations Portal → Knowledge Bases → Create. Give it a name and description.
  2. Add documents — upload files directly, or configure an external source to sync automatically.
  3. Wire into a flow — add a RAG Query or KB Search node to your bot's flow and select the KB.
  4. Test — use the test panel to run real questions and verify retrieval quality.

Sections ​

  • Documents — supported formats, uploading, metadata, updating
  • External sources — automatic sync from SharePoint, Google Drive, Confluence, ServiceNow, and 15+ other providers
  • RAG settings — tuning chunk size, top K, similarity threshold, embedding models

Multi-KB strategies ​

You can create multiple KBs per tenant and reference them from different bots (or the same bot on different pathways):

  • One KB per topic — Product FAQ, Billing, Policies. RAG Query selects the KB based on the flow's route.
  • One KB per audience — Customer, Partner, Employee. Filter by metadata (see Documents) instead of separate KBs.
  • One KB per language — En, Es, Fr. Retrieve based on sys_locale.

Practice ​

  • Start small — a focused 20–100-document KB retrieves better than a 10,000-document catch-all. Add content as your bot proves it needs it.
  • Iterate on retrieval quality — the test panel with Debug Mode on shows exactly which chunks matched. Bad matches usually mean bad chunk size, poor document formatting, or missing content.
  • Update on schedule — external sources sync automatically; uploaded documents need re-uploading if the source changes.

AdaptCX — product documentation

© AdaptCX