Skip to main content
Configure knowledge sources, embedding, chunking, and retrieval settings for RAG-powered agents.

​
Quick Start

1

Simple Usage

Add knowledge sources directly:
from praisonaiagents import Agent

agent = Agent(
    name="Knowledge Agent",
    instructions="Answer questions from the documentation",
    knowledge=["docs/", "data.pdf"]
)
2

With Configuration

Configure retrieval settings:
from praisonaiagents import Agent
from praisonaiagents.config import KnowledgeConfig

agent = Agent(
    name="Knowledge Agent",
    instructions="Answer questions from the documentation",
    knowledge=KnowledgeConfig(
        sources=["docs/"],
        embedder="openai",
        chunking_strategy="semantic",
        retrieval_k=5,
        rerank=True
    )
)

​
Configuration Options

from praisonaiagents.config import KnowledgeConfig, ChunkingStrategy

config = KnowledgeConfig(
    # Knowledge sources
    sources=["docs/", "data.pdf"],
    
    # Embedder configuration
    embedder="openai",
    embedder_config=None,
    
    # Chunking settings
    chunking_strategy=ChunkingStrategy.SEMANTIC,
    chunk_size=1000,
    chunk_overlap=200,
    chunker=None,  # Alternative dict config
    
    # Retrieval settings
    retrieval_k=5,
    retrieval_threshold=0.0,
    
    # Reranking
    rerank=False,
    rerank_model=None,
    
    # Auto-retrieval
    auto_retrieve=True,
    
    # Vector store
    vector_store=None,
    
    # Advanced config
    config=None
)
ParameterTypeDefaultDescription
sourcesList[str][]Files, directories, or URLs
embedderstr"openai"Embedding provider
embedder_configDict | NoneNoneEmbedder-specific settings
chunking_strategystr | ChunkingStrategy"semantic"Chunking method (fixed, semantic, sentence, paragraph)
chunk_sizeint1000Target chunk size in tokens
chunk_overlapint200Overlap between chunks
chunkerDict | NoneNoneAlternative chunker config dict
retrieval_kint5Number of chunks to retrieve
retrieval_thresholdfloat0.0Minimum similarity threshold
rerankboolFalseEnable result reranking
rerank_modelstr | NoneNoneModel for reranking
auto_retrieveboolTrueAuto-inject relevant context
vector_storeDict | NoneNoneVector database settings
configDict | NoneNoneAdvanced configuration dict

​
Common Patterns

​
Pattern 1: High-Precision Retrieval

from praisonaiagents import Agent
from praisonaiagents.config import KnowledgeConfig

agent = Agent(
    name="Precise Agent",
    instructions="Provide accurate answers",
    knowledge=KnowledgeConfig(
        sources=["docs/"],
        retrieval_k=10,
        rerank=True,
        retrieval_threshold=0.7
    )
)

​
Pattern 2: Custom Chunking

from praisonaiagents import Agent
from praisonaiagents.config import KnowledgeConfig

agent = Agent(
    name="Custom Chunk Agent",
    instructions="Process large documents",
    knowledge=KnowledgeConfig(
        sources=["large_docs/"],
        chunker={
            "type": "semantic",
            "chunk_size": 512,
            "chunk_overlap": 100
        }
    )
)

​
Pattern 3: External Vector Store

from praisonaiagents import Agent
from praisonaiagents.config import KnowledgeConfig

agent = Agent(
    name="Qdrant Agent",
    instructions="Use external vector database",
    knowledge=KnowledgeConfig(
        sources=["docs/"],
        vector_store={
            "provider": "qdrant",
            "url": "http://localhost:6333",
            "collection": "my_docs"
        }
    )
)

​
Best Practices

Semantic chunking preserves context better than fixed-size chunks, especially for technical documentation.
Reranking improves retrieval precision by using a cross-encoder to reorder results.
Use retrieval_threshold to filter low-quality matches and prevent hallucination.
Keep auto_retrieve=True to automatically inject relevant context into prompts.

Knowledge Overview

Learn about the knowledge system

RAG

RAG configuration and features