Knowledge Sources & RAG

Knowledge in CrewAI lets you attach external documents or data to agents or crews. During task execution, the runtime indexes that content and retrieves relevant chunks when the LLM needs context—this pattern is RAG (Retrieval Augmented Generation): the model answers using retrieved passages, not only its training data.

Knowledge source types

CrewAI ships with several built-in sources. File-based sources typically live under a knowledge folder at your project root; use paths relative to that directory.

Crew-level knowledge

Pass knowledge_sources on the Crew. Every agent in that crew can query the same shared knowledge (along with any agent-specific sources you add later). The sketch below assumes support_agent and answer_task are already defined.

from crewai import Agent, Crew, Task
from crewai.knowledge.source.string_knowledge_source import StringKnowledgeSource
 
company_info = StringKnowledgeSource(
    content="Our company was founded in 2020. We specialize in AI solutions for healthcare...",
)
 
crew = Crew(
    agents=[support_agent],
    tasks=[answer_task],
    knowledge_sources=[company_info],
)

Agent-level knowledge

Pass knowledge_sources on an Agent when only that role should see certain material. Those sources are scoped to that agent’s retrieval index. Here product_docs is any knowledge source instance (for example StringKnowledgeSource or a file-based source).

agent = Agent(
    role="Product Expert",
    goal="Answer product questions accurately",
    backstory="...",
    knowledge_sources=[product_docs],
)

Crew knowledge vs agent knowledge

• Crew knowledge is shared: all agents in the crew can retrieve from those sources.
• Agent knowledge is private to that agent: other agents do not automatically get it.

An agent with both crew-wide and agent-only sources effectively searches crew knowledge plus its own agent sources.

Fine-tuning retrieval with KnowledgeConfig

Tune how many chunks are returned and how strict similarity filtering is using KnowledgeConfig. Pass it via knowledge_config on an Agent or Crew (depending on your setup).

from crewai.knowledge.knowledge_config import KnowledgeConfig
 
knowledge_config = KnowledgeConfig(results_limit=10, score_threshold=0.5)
 
agent = Agent(
    # ...
    knowledge_config=knowledge_config,
)

• results_limit: maximum number of relevant chunks to return (default is often around 3 in recent versions; check your CrewAI release notes).
• score_threshold: minimum similarity score for a chunk to be included (higher = stricter).

Embedder configuration

Knowledge is embedded for vector search. By default CrewAI often uses OpenAI embeddings (for example text-embedding-3-small), which requires an API key even if your chat LLM is another provider. You can set a different embedder per agent or crew—for example OpenAI, Ollama, or Azure OpenAI—using a dict with provider and config (model name, endpoint, and so on per provider docs).

agent = Agent(
    # ...
    embedder={"provider": "openai", "config": {"model": "text-embedding-3-small"}},
)

Agents without their own embedder can fall back to the crew’s embedder when one is configured.

Key takeaways

• Knowledge gives agents grounded, retrievable context (RAG) instead of stuffing everything into prompts.
• Choose crew sources for policies and facts everyone needs; use agent sources for role-specific manuals or data.
• KnowledgeConfig controls how much and how selective retrieval is via results_limit and score_threshold.
• Embedders are configurable per provider; align embedding setup with your LLM and compliance constraints.