Vector Store

Description

The Vector-Store component allows interaction with a vector store. It is available if at least one LLM integration (MistralAI, OpenAI) is configured in the Admin Panel console (Integration tab). It allows users to add, search and truncate the vector store based on the Operation parameter:

• The add operation uses the selected LLM embedding model to convert provided text document ("content") to the embedding vector and stores the text and the vector in the vector store.
• The search operation searches for text documents which are close to the provided text.
• The truncate operation removes all previously added embedding vectors and associated texts.

Parameters and configuration

Name	Description
Vector store	The vector store instance to use; configure Qdrant or Weaviate integration to use this component. See also Additional considerations section for additional information.
Operation	Operation to perform.
LLM Integration	add and search operations only - LLM integration to use
Collection	Some vector stores allow dividing its space into separate collections. This parameter is available only if selected vector store supports collections.
Embedding model name	add and search operations only - LLM provided model to use when adding or searching embeddings.
Content	add and search operations only - text document to use as the basis of the add or search.
Max results	search operation only - upper boundary of number of text documents to return.
Metadata	add operation - metadata to be put in the vector store together with vector and text document. Provided metadata will be available in results returned from the vector store after performing the search operation. Metadata must be in a form of string to string map.
Output variable name	Variable name under which node result will be available in a subsequent nodes.

Returned value

The return value will be stored in the user defined variable (see above) and depends on the operation performed

Operation	Returned value
add	id of the stored embedding
search	List of records, each containing: - "content" - the text document which was stored as an embedding vector - "metadata" - the metadata associated with the stored document - "score" - the similarity score indicating how close the result is to the query
truncate	For this operation the value is not deterministic.

Additional considerations

• add/truncate operations are also performed when testing scenarios on the selected vector store.
• For testing purposes the “in-memory” vector store is provided. The "in-memory" vector store is not durable or accessible outside the Nussknacker Cloud and therefore should not be used for scenarios running in production.
• The vector length used by the vector store collection must be of the same length as the vector length produced by embedding model (e.g., 1024 for mistral-embed model).
• Nussknacker does not create collections in a vector store. If needed it must be created before using it from a scenario.