Knowledge retrieval Node

Overview

The Knowledge Retrieval Node searches and retrieves relevant documents from your knowledge collections based on a query. It supports semantic search through vector embeddings and allows filtering of results using vector store-specific filtering syntax. Each knowledge collection can be configured with different vector stores (Milvus, Qdrant, Chroma, Pinecone, ...).

Usage cost: 0.1 credit / document

Configuration

Settings

1. Collection Selection

   • Collection*: Select the knowledge collection to search

   • Number of Documents: Number of results to return (default: 5)

   • Query*: Search query used to find relevant documents

   • Filtering: Optional filtering condition based on metadata

Output Ports

• knowledges (Document[]): Retrieved documents with full metadata

• documents_content (string[]): Array of document contents only

Filtering example depending on vector store

Milvus/Zilliz

metadata["field"] == "value"

Qdrant

{
  "must": [
    {"key": "metadata.field", "match": {"value": "exact_match"}}
  ]
}

Chroma

{
  "field": {"$eq": "value"}
}

Pinecone

{
  "field": {"$eq": "value"}
}

Best Practices

1. Query Construction

   • Keep queries clear and focused

   • Include key terms relevant to desired content

   • Consider semantic meaning rather than exact keywords

   • Test queries with different phrasings

2. Result Optimization

   • Adjust number of documents based on use case

   • Use filtering to narrow down results, or to built multi-tenant data retrieval

   • Consider document length when setting limits

3. Filtering Usage

   • Follow vector store-specific syntax

   • Test filters with sample data

   • Use appropriate operators for data types

For vector store specific filter syntax, refer to:

• Milvus: Filtering Documentation

• Zilliz: Filtering Documentation

• Qdrant: Filtering Documentation

• Chroma: Where Filters Documentation

• Pinecone: Metadata Filtering Guide