Retrieve context
Retrieve the most relevant documents from a vector store based on a user query.
Context retrieval is the step where semantic search happens: a query is embedded and compared against stored document embeddings to find the most similar content.
When this is useful
Retrieve context when you want to:
- Find documents related to a user question
- Ground model responses in external knowledge
- Prepare context for retrieval-augmented generation (RAG)
If you don’t yet have indexed documents, start with Create a vector store.
Retrieve context in KNIME
Prerequisites
- A vector store (FAISS or Chroma)
- The same embedding model used to create the store
- A query provided as text (string column)
Retrieving context consists of three steps:
- Load the vector store
- Provide a query
- Retrieve the most similar entries
Step 1: Load the vector store
If the vector store was created in the same workflow, you can pass it directly to the retriever.
If the store was saved earlier, load it using the Model Reader node. For details on saving and loading vector stores, see Reuse a vector store.
Step 2: Provide a query
Provide the query you want to search for in the vector store. This is simply a string value and can come from any upstream node.
Step 3: Retrieve relevant context
Use the Vector Store Retriever node to perform the search.
The Vector Store Retriever performs two operations:
- Creates an embedding of the query
- Finds the most similar results from the vector store
To create the query embedding, the node needs the same embedding model that was used to embed the documents in the vector store. If a remote embedding model was used, this means you need:
- an active internet connection
- valid credentials to access the model
Configure the node to:
- use the same embedding model as the vector store
- embed the query automatically
- return the top k most similar documents
The node compares the query embedding against the stored vectors and returns the most relevant entries.
Result
The output table contains the retrieved documents, ordered by semantic similarity.
Each row includes:
- the original document text
- a similarity score or distance