Vector Collections

A vector collection is a set of documents that have been transformed into vector embeddings for fast semantic similarity search. Each collection lives under a .gbkb folder and is identified by a unique name.

Creating a Collection

Use the ADD_KB keyword in a dialog script:

ADD_KB "company-policies"

This creates a new collection named company-policies in the bot’s knowledge base.

Adding Documents

Documents can be added directly from files or by crawling a website:

ADD_KB "company-policies"   ' adds a new empty collection
ADD_WEBSITE "https://example.com/policies"

The system will download the content, split it into chunks, generate embeddings using the default LLM model, and store them in the collection.

Managing Collections

SET_KB "collection-name" – selects the active collection for subsequent ADD_KB or FIND calls.
LIST_KB – (not a keyword, but you can query via API) lists all collections.

Use in Dialogs

When a collection is active, the FIND keyword searches across its documents, and the GET_BOT_MEMORY keyword can retrieve relevant snippets to inject into LLM prompts.

SET_KB "company-policies"
FIND "vacation policy" INTO RESULT
TALK RESULT

Technical Details

Embeddings are generated with the BGE‑small‑en‑v1.5 model.
Vectors are stored in VectorDB (see Chapter 04).
Each document is chunked into 500‑token pieces for efficient retrieval.

1.5 KiB Raw Blame History Unescape Escape