Knowledge Retrieval

# Knowledge Retrieval

Retrieve relevant knowledge segments from a knowledge base. Supports three search methods: vector search (semantic similarity), keyword search (full-text matching), and hybrid search (combining both). Optionally, AI summarization can be enabled to generate a concise answer from the retrieved documents.

# Creating a Knowledge Base

Before using Knowledge Retrieval, you need to create a knowledge base in the Knowledge section. The following configuration is required:

Name: The name of the knowledge base.
Description: A description of the knowledge base.
File: Upload a document file to be indexed. Supported formats include text files.
LLM Credential: Select an LLM API Key credential. The embedding model from this provider will be used to vectorize the document content. You can create an LLM API Key credential in the Credential section.
Vector Database: Select the vector database for storing embeddings. The built-in vector database (builtin-vec-db) is available by default.
Chunk Size: The maximum size (in characters) of each document chunk. Default is 1024.
Chunk Overlap: The number of overlapping characters between adjacent chunks. Default is 50.
Separator: The separator used to split the document into chunks. Default is \n\n (double newline).
Top K: The default number of segments returned during retrieval.
Score Threshold: The minimum relevance score threshold for retrieval results.

When a knowledge base is created, the system will:

Split the uploaded document into chunks based on the configured chunk size, overlap, and separator.
Generate embedding vectors for each chunk using the selected LLM provider's embedding model.
Store the vectors in the vector database for semantic search.
Build a full-text index for each chunk to support keyword search.

You can view the chunked segments of a knowledge base at any time, and update or delete the knowledge base as needed.

# Input

Knowledge (Required): Select the knowledge base to search from. Knowledge bases are created and managed in the Knowledge section, where you can upload documents and configure chunking settings.
Input (Required): The query text to search for in the knowledge base. You can use variable expressions to integrate outputs from other app nodes.

Search Method: The search algorithm to use. Default is vector.

Method	Description
vector	Semantic vector search — finds documents with similar meaning using embedding models. Best for natural language queries.
keyword	Full-text keyword search — finds documents containing the query keywords. Best for exact term matching.
hybrid	Combines vector and keyword search with weighted scoring (vector: 0.7, keyword: 0.3). Best for comprehensive results.

Top K: The maximum number of knowledge segments to return. Default is based on knowledge base settings.
LLM Model: Select an LLM model for AI summarization. When AI summarization is enabled, the retrieved documents will be summarized by this model into a concise answer.
AI Summarize: Whether to enable AI summarization. When enabled, the LLM model will generate a summary answer based on the retrieved knowledge segments.

# Output

Returns an object containing the AI-generated answer (if summarization is enabled) and the list of retrieved document segments:

{
  "answer": "The capital of France is Paris, which is known for the Eiffel Tower.",
  "docs": [
    {
      "segmentId": 1,
      "score": 0.95,
      "pageContent": "Paris is the capital of France. It is famous for the Eiffel Tower..."
    },
    {
      "segmentId": 5,
      "score": 0.82,
      "pageContent": "France is a country in Western Europe. Its capital Paris..."
    }
  ]
}

Field	Description
answer	AI-generated summary answer (empty if AI summarization is disabled)
docs	Array of retrieved knowledge segments
docs[].segmentId	The segment ID within the knowledge base
docs[].score	The relevance score (0–1, higher is more relevant)
docs[].pageContent	The text content of the knowledge segment

# How It Works

# Vector Search

The input query is converted into an embedding vector using the configured LLM provider's embedding model.
The vector is compared against pre-indexed document vectors in the knowledge base.
The top K most similar segments are returned ranked by cosine similarity score.

# Keyword Search

The input query is tokenized using a Unicode-aware tokenizer (supports Chinese and English).
Full-text search is performed against the indexed document segments.
The top K matching segments are returned ranked by relevance score.

# Hybrid Search

Both vector search and keyword search are executed in parallel (each retrieves 2×TopK results).
Results are merged by segment ID.
A weighted final score is calculated: 0.7 × vectorScore + 0.3 × keywordScore.
Results are sorted by final score and the top K segments are returned.

← Database Query →