The semantic search nodes embeds and queries documents semantically at run time. In contrast to the knowledge base, which is a permanent store of data, documents embedded using the semantic search node are immediately deleted after the pipeline finishes running.

​
Node Inputs

  1. Search Query: The query that will be used to search the embedded documents semantically for relevant pieces
    • Type: Text
  2. Documents for Search: The text that will be semantically searched
    • Type: Text

​
Node Parameters

On the face of the node:

  1. Embedding model: The embedding model to use to embed the documents
    • Type: Dropdown

In the gear:

  1. Max chunks per query: the maximum number of pieces of data to be returned per query
  2. Enable Filter: enable the ability to write a metadata filter
  3. Enable Context: enable a text field to provide additional context for the search query
  4. Re-rank documents: Performs an additional reranking step to reorder the documents by relevance to the query
  5. Score Cutoff: The minimum relevancy score (between 0 and 1) that each piece of data will have semantically to the query
  6. Retrieval Unit: Return the most relevant chunks (text content) or Documents (will return document metadata)
  7. Transform Query: Transform the query for better results
  8. Answer Multiple Questions: Extract separate questions from the query and retrieve content separately for each question to improve search performance
  9. Expand Query: Expand query to improve semantic search
  10. Do Advanced QA: Use additional LLM calls to analyze each document to improve answer correctness
  11. Show Immediate Steps: Display the process the knowledge base is conducting at a given time in the chatUI
  12. Format Context for an LLM: Do an additional LLM call to format output

​
Node Outputs

If Retrieval Unit is set to Chunks

  1. Chunks: Semantically similar chunks retrieved from the documents
    • Type: List<Text>
    • Example usage: {{semantic_search_0.chunks}}

If Retrieval Unit is set to Documents

  1. Documents: Metadata for semantically similar documents retrieved from the documents
    • Type: List<Text>
    • Example usage: {{semantic_search_0.documents}}

If “Do Advanced QA” is enabled

  1. Response: A direct answer to the query
    • Type: Text
    • Example usage: {{semantic_search_0.response}}

​
Considerations

  • Use a semantic search node when your pipeline loads new data at run time for querying. Use a Knowledge Base Reader node when you want to query previously loaded data that has already been loaded.
  • If the semantic search node is not returning relevant information to a query, try increasing the number of max chunks per query in the gear of the knowledge base.
  • For debugging purposes, you may attach an output node to the semantic search node to view the chunks that are returned for the query.

​
Example

The below example is a pipeline which answers questions about on the contents of a website (example is with VectorShift’s website).

  • Input Node: The query about the website.
  • Text Node: The URL for the website.
  • Scrape URL node: Scrape the website URL
    • URL: {{text_0.text}}
  • Semantic Search Node: Find semantically similar chunks to a question (input node) based on a website (Scrape URL node).
    • Search Query: {{input_0.text}}
    • Documents for Search: {{url_loader_0.content}}
  • LLM Node: Answer a user query (input node) based on relevant information (semantic search node).
    • Prompt: {{input_0.text}} and {{semantic_search_0.chunks}}
  • Output Node: Display the LLM’s response.
    • Output: {{openai_0.response}}