Skip to main content

​
Overview

The Search Knowledge Node performs retrieval-augmented generation (RAG) by searching against one or more knowledge bases. It takes a query expression, executes vector similarity search (optionally combined with BM25 keyword search), and returns the most relevant document chunks along with their scores and metadata. This node is a core building block for any workflow that needs to ground LLM responses in your organization’s documents and data.

​
How It Works

1

Resolve Query

The node evaluates the query expression, which can reference workflow context variables such as the user’s message or a transformed version of it.
2

Search Knowledge Bases

The query is sent to each configured knowledge base. Depending on the knowledge base configuration, this may be a dense vector search, a BM25 keyword search, or a hybrid combination of both.
3

Score and Filter

Results are ranked by relevance score. Chunks below the score_threshold are discarded, and only the top top_k results are retained.
4

Output Results

The retrieved chunks, along with their scores and source metadata, are written to the workflow context for downstream nodes to consume.

​
Configuration

{
  "type": "search-knowledge-node",
  "config": {
    "knowledge_base_ids": [
      "kb_01HX7A9B2C3D4E5F6G7H8J9K",
      "kb_01HX7B8C3D4E5F6G7H8J9K0L"
    ],
    "query": "{{user_message}}",
    "top_k": 5,
    "score_threshold": 0.7,
    "search_mode": "hybrid",
    "output_variable": "search_results"
  }
}
ParameterTypeDefaultDescription
knowledge_base_idsstring[]—IDs of the knowledge bases to search (required)
querystring—Query expression, supports variable references like {{user_message}}
top_knumber5Maximum number of chunks to return
score_thresholdnumber0.0Minimum relevance score (0.0 to 1.0) to include a chunk
search_modestring"vector"Search strategy: "vector", "keyword", or "hybrid"
output_variablestring"search_results"Context variable name for the output

​
Search Modes

ModeDescriptionBest For
vectorDense embedding similarity search using pgvectorSemantic queries where meaning matters more than exact terms
keywordBM25 keyword-based searchExact term matching, code identifiers, proper nouns
hybridCombines vector and keyword results with reciprocal rank fusionGeneral-purpose retrieval with the best balance of recall and precision
Hybrid search is recommended for most use cases. It catches both semantic matches that keyword search would miss and exact matches that vector search might overlook.

​
Output Structure

The node writes an array of retrieved chunks to the workflow context: 
{
  "search_results": [
    {
      "content": "The refund policy allows returns within 30 days of purchase...",
      "score": 0.92,
      "metadata": {
        "source": "refund-policy.pdf",
        "page": 3,
        "knowledge_base_id": "kb_01HX7A9B2C3D4E5F6G7H8J9K",
        "chunk_index": 12
      }
    },
    {
      "content": "Exceptions to the return policy include digital products...",
      "score": 0.85,
      "metadata": {
        "source": "refund-policy.pdf",
        "page": 4,
        "knowledge_base_id": "kb_01HX7A9B2C3D4E5F6G7H8J9K",
        "chunk_index": 15
      }
    }
  ]
}
FieldDescription
contentThe text content of the retrieved chunk
scoreRelevance score between 0.0 and 1.0
metadata.sourceOriginal document filename
metadata.pagePage number in the source document (if applicable)
metadata.knowledge_base_idWhich knowledge base the chunk came from
metadata.chunk_indexPosition of the chunk within the document

​
Integration with Reranker Node

For improved retrieval quality, chain the Search Knowledge Node with a Reranker Node. The reranker uses a cross-encoder model to re-score the initial results, often surfacing more relevant chunks that vector search alone might rank lower. Recommended pattern:
  1. Retrieve a larger initial set (top_k: 20) from the Search Knowledge Node
  2. Pass the results through the Reranker Node to select the best 5
  3. Feed the reranked results to the AI Agent Node as context
{
  "type": "reranker-node",
  "config": {
    "input_variable": "search_results",
    "top_k": 5,
    "model": "cross-encoder",
    "output_variable": "reranked_results"
  }
}

​
Searching Multiple Knowledge Bases

When multiple knowledge_base_ids are specified, the node searches all of them and merges the results into a single ranked list. This is useful when your content spans multiple knowledge bases (e.g., product docs + support tickets + internal wiki). Results from different knowledge bases are interleaved by score, so the most relevant chunks appear first regardless of their source.

​
Example: Customer Support RAG

A complete RAG workflow for answering customer questions: 
{
  "nodes": [
    {
      "id": "start",
      "type": "start-node"
    },
    {
      "id": "search",
      "type": "search-knowledge-node",
      "config": {
        "knowledge_base_ids": ["kb_product_docs", "kb_faq"],
        "query": "{{user_message}}",
        "top_k": 10,
        "score_threshold": 0.6,
        "search_mode": "hybrid"
      }
    },
    {
      "id": "rerank",
      "type": "reranker-node",
      "config": {
        "input_variable": "search_results",
        "top_k": 3
      }
    },
    {
      "id": "agent",
      "type": "ai-agent-node",
      "config": {
        "agent_mode": "standard",
        "model": "gpt-4o",
        "system_prompt": "Answer the user's question based on the provided context. If the context does not contain relevant information, say so."
      }
    },
    {
      "id": "end",
      "type": "end-node"
    }
  ]
}

​
Best Practices

A score_threshold of 0.6-0.7 works well for most use cases. Too low (below 0.5) may include irrelevant results that confuse the LLM. Too high (above 0.9) may filter out useful results.
Unless you have a specific reason to use pure vector or keyword search, hybrid mode provides the best recall across different query types.
Retrieve more results than you need (top_k: 15-20) and use a Reranker Node to select the best subset. This two-stage approach significantly improves retrieval quality.
For complex user messages, use a Variable Node or AI Agent Node upstream to extract or rephrase the search query before passing it to the Search Knowledge Node.

​
Next Steps

Knowledge Base

Set up and manage knowledge bases for your workflows

Document Extract

Process uploaded documents within workflows

Database Node

Query structured data directly from your workflows

AI Agent Node

Use retrieved knowledge as context for LLM responses