Skip to main content
Knowledge bases provide agents with access to domain-specific information, documents, and organizational knowledge through semantic search.

​
Available Knowledge Base Types

Milvus

Vector Database ✅Production-ready vector database with semantic search capabilities.
Additional vector database integrations will be supported in future releases.

​
Milvus Knowledge Base

Milvus is a vector database that enables semantic search over your documents using dense vector embeddings. Knowledge bases support customizable chunking strategies, embedding models, and structured data processing.

​
Basic Configuration

agents:
  - name: "support_agent"
    agent_type: "llm_agent"
    
    knowledge_bases:
      - name: "company_docs"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_abc123"
        retrieval_config:
          top_k: 10
          search_ef: 64
          metric_type: "COSINE"

​
Creating with API

Knowledge bases are created via the API with full control over processing configuration. See Create Knowledge Base API for complete documentation.

​
Required Fields

​
name
string
required
Unique identifier for this knowledge base instance
​
knowledge_base_type
string
required
Must be set to "milvus"
​
knowledge_base_id
string
required
Knowledge base ID for isolation and multi-tenancy

​
Retrieval Configuration

The retrieval_config object controls how documents are retrieved:
​
retrieval_config.top_k
integer
default:"10"
Number of most relevant results to returnRange: 1-1000
​
retrieval_config.search_ef
integer
default:"64"
HNSW search parameter controlling accuracy vs speed trade-offRange: 1-512
Higher values = more accurate but slower
​
retrieval_config.metric_type
string
default:"COSINE"
Distance metric for similarity searchOptions:
  • COSINE - Cosine similarity (recommended)
  • L2 - Euclidean distance
  • IP - Inner product
​
retrieval_config.offset
integer
default:"0"
Number of results to skip (for pagination)
​
retrieval_config.score_threshold
float
Minimum similarity score thresholdRange: 0.0-1.0
Only return results above this score
​
retrieval_config.strategy
string
default:"dense"
Search strategy for retrievalOptions:
  • dense - Vector-based semantic search only (default, fastest)
  • hybrid - Combines dense vector + BM25 keyword search
  • bm25 - BM25 keyword-based search only
  • hybrid_rrf - Hybrid with Reciprocal Rank Fusion for improved ranking

​
Usage Examples

​
Basic Knowledge Base Agent

agents:
  - name: "support_agent"
    agent_type: "llm_agent"
    
    knowledge_bases:
      - name: "support_kb"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_support_001"
        retrieval_config:
          top_k: 5
          metric_type: "COSINE"
    
    system_prompt: |
      You are a customer support agent with access to our support documentation.
      
      Always search the knowledge base first for answers to customer questions.
      Provide accurate information based on our documentation.

​
Hybrid Search with RRF Fusion

agents:
  - name: "research_agent"
    agent_type: "react_agent"
    
    tools:
      - "tavily_search"
    
    knowledge_bases:
      - name: "research_kb"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_research_001"
        retrieval_config:
          strategy: "hybrid_rrf"  # Combines dense + BM25 with RRF
          top_k: 20
          search_ef: 128
          metric_type: "COSINE"
          score_threshold: 0.7
    
    system_prompt: |
      You are a research assistant with access to internal research papers.
      
      Search the knowledge base for relevant research before using web search.
      Prioritize high-quality, relevant results.

​
Query Expansion

agents:
  - name: "advanced_support_agent"
    agent_type: "llm_agent"
    
    knowledge_bases:
      - name: "product_kb"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_product_001"
        retrieval_config:
          strategy: "dense"
          top_k: 15
          search_ef: 96
          metric_type: "COSINE"
          
          # Query expansion for better recall
          query_model: "gpt-4o"
          max_num_query_expansions: 3
    
    system_prompt: |
      You are an expert product support agent.
      Use query expansion to find relevant documentation.

​
Multiple Knowledge Bases

agents:
  - name: "comprehensive_agent"
    agent_type: "llm_agent"
    
    knowledge_bases:
      - name: "technical_docs"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_tech_001"
        retrieval_config:
          top_k: 10
      
      - name: "company_policies"
        knowledge_base_type: "milvus"
        knowledge_base_id: "kb_policy_001"
        retrieval_config:
          top_k: 5
    
    system_prompt: |
      You have access to multiple knowledge bases:
      - Technical documentation
      - Company policies
      
      Search the appropriate knowledge base based on the question type.

​
Chunking Strategies

Knowledge bases support three chunking strategies configured during creation:

​
Recursive (Default)

Splits text using multiple separators in order (paragraphs → sentences → words). Best for general documents. Parameters:
  • strategy: "recursive"
  • chunk_size: Size in words/characters (default: 500)
  • chunk_overlap: Overlap between chunks (default: 50)
  • split_by: "word" | "char" (only these two options)
  • recursive_separators: Array of separators to try in order
Example: 
{
  "strategy": "recursive",
  "chunk_size": 500,
  "chunk_overlap": 50,
  "split_by": "word",
  "recursive_separators": ["\n\n", "\n", ". ", " "]
}

​
Hierarchical

Creates multi-level chunks preserving document structure. Ideal for academic papers and structured documents. Parameters:
  • strategy: "hierarchical"
  • hierarchical_block_sizes: Descending array of block sizes (e.g., [700, 350, 150])
  • chunk_overlap: Overlap between chunks
  • split_by: "word" | "sentence"
Example: 
{
  "strategy": "hierarchical",
  "chunk_overlap": 70,
  "split_by": "word",
  "hierarchical_block_sizes": [700, 350, 150]
}

​
Fixed

Simple fixed-size chunks. Fastest processing for straightforward documents. Parameters:
  • strategy: "fixed"
  • chunk_size: Fixed chunk size
  • chunk_overlap: Overlap between chunks
  • split_by: "word" | "sentence"
Example: 
{
  "strategy": "fixed",
  "chunk_size": 500,
  "chunk_overlap": 50,
  "split_by": "word"
}

​
Structured Data Processing

CSV and Excel files (.csv, .xlsx, .xls) support specialized processing:

​
Configuration

{
  "document_config": {
    "strategy": "recursive",
    "chunk_size": 500,
    "chunk_overlap": 50,
    "split_by": "word"
  },
  "structured_config": {
    "rows_per_batch": 10,
    "table_format": "csv",
    "csv_content_column": "text",
    "csv_conversion_mode": "row"
  }
}

​
Structured Config Parameters

​
rows_per_batch
integer
default:"10"
Number of rows to combine into one searchable chunk (1-20)Lower values = More precise retrieval, slower ingestion
Higher values = Faster ingestion, broader context
​
table_format
string
default:"csv"
Output format for table dataOptions:
  • csv - Comma-separated values
  • markdown - Markdown table format
​
csv_content_column
string
default:"text"
Column name containing the main text content (CSV only)Required for row mode processing
​
csv_conversion_mode
string
default:"row"
Processing mode for CSV filesOptions:
  • row - One document per row (precise retrieval)
  • file - One document per file (holistic context)

​
Embedding Providers

Choose your embedding model during knowledge base creation:

​
Azure OpenAI

{
  "provider": "azure_openai",
  "model": "text-embedding-ada-002",
  "dimensions": 1536
}
Available model:
  • text-embedding-ada-002 (1536 dimensions)

​
Mistral AI

{
  "provider": "mistral",
  "model": "mistral-embed",
  "dimensions": 1024
}
Available model:
  • mistral-embed (1024 dimensions)

​
Telekom OTC

{
  "provider": "telekom_otc",
  "model": "text-embedding-bge-m3",
  "dimensions": 1024
}
Available models:
  • text-embedding-bge-m3 (1024 dimensions) - BGE multilingual model
  • jina-embeddings-v2-base-de (768 dimensions) - German-optimized
  • jina-embeddings-v2-base-code (768 dimensions) - Code-optimized
  • tsi-embedding-colqwen2-2b-v1 (1024 dimensions) - TSI ColQwen2

​
Best Practices

​
Chunking

General Documents: Use recursive strategy with 500 word chunks and 50 word overlap.
Code Files: Use smaller chunks (300 words) with recursive separators ["\n\n", "\n", " "].
Research Papers: Use hierarchical strategy with block sizes [700, 350, 150] to preserve structure.
CSV/Excel Data: Start with rows_per_batch: 10 and adjust based on row size and retrieval precision needs.
Chunk overlap must be less than chunk size (or smallest block size for hierarchical). The API will reject invalid configurations.

​
Retrieval

Top K Selection: Start with top_k: 5-10 for most use cases. Increase if you need more context.
Search Accuracy: Use search_ef: 64 for balanced performance. Increase to 128+ for higher accuracy needs.
Metric Type: Use COSINE for most semantic search applications as it’s normalized and works well with embeddings.
Score Threshold: Set a score_threshold (e.g., 0.7) to filter out low-relevance results.
Higher search_ef values improve accuracy but increase query latency. Balance based on your performance requirements.

​
Next Steps

Agent Capabilities

Explore all agent capabilities including tools and streaming

Agent Types

Learn about different agent types