Knowledge

Knowledge is domain-specific information that the Agent can search at runtime to make better decisions (dynamic few-shot learning) and provide accurate responses (agentic RAG). Knowledge is stored in a vector db and this searching on demand pattern is called Agentic RAG.

Dynamic Few-Shot Learning: Text2Sql Agent

Example: If we’re building a Text2Sql Agent, we’ll need to give the table schemas, column names, data types, example queries, common “gotchas” to help it generate the best-possible SQL query.

We’re obviously not going to put this all in the system prompt, instead we store this information in a vector database and let the Agent query it at runtime.

Using this information, the Agent can then generate the best-possible SQL query. This is called dynamic few-shot learning.

Agno Agents use Agentic RAG by default, meaning when we provide knowledge to an Agent, it will search this knowledge base, at runtime, for the specific information it needs to achieve its task.

The pseudo steps for adding knowledge to an Agent are:

from agno.agent import Agent, AgentKnowledge

# Create a knowledge base for the Agent
knowledge_base = AgentKnowledge(vector_db=...)

# Add information to the knowledge base
knowledge_base.load_text("The sky is blue")

# Add the knowledge base to the Agent and
# give it a tool to search the knowledge base as needed
agent = Agent(knowledge=knowledge_base, search_knowledge=True)

We can give our agent access to the knowledge base in the following ways:

• We can set search_knowledge=True to add a search_knowledge_base() tool to the Agent. search_knowledge is True by default if you add knowledge to an Agent.
• We can set add_references=True to automatically add references from the knowledge base to the Agent’s prompt. This is the traditional 2023 RAG approach.

If you need complete control over the knowledge base search, you can pass your own retriever function with the following signature:

def retriever(agent: Agent, query: str, num_documents: Optional[int], **kwargs) -> Optional[list[dict]]:
  ...

This function is called during search_knowledge_base() and is used by the Agent to retrieve references from the knowledge base.

Vector Databases

While any type of storage can act as a knowledge base, vector databases offer the best solution for retrieving relevant results from dense information quickly. Here’s how vector databases are used with Agents:

1. Chunk the information

   Break down the knowledge into smaller chunks to ensure our search query returns only relevant results.

2. Load the knowledge base

   Convert the chunks into embedding vectors and store them in a vector database.

3. Search the knowledge base

   When the user sends a message, we convert the input message into an embedding and “search” for nearest neighbors in the vector database.

Example: RAG Agent with a PDF Knowledge Base

Let’s build a RAG Agent that answers questions from a PDF.

Step 1: Run PgVector

Let’s use PgVector as our vector db as it can also provide storage for our Agents.

Install docker desktop and run PgVector on port 5532 using:

docker run -d \
  -e POSTGRES_DB=ai \
  -e POSTGRES_USER=ai \
  -e POSTGRES_PASSWORD=ai \
  -e PGDATA=/var/lib/postgresql/data/pgdata \
  -v pgvolume:/var/lib/postgresql/data \
  -p 5532:5432 \
  --name pgvector \
  agnohq/pgvector:16

Step 2: Traditional RAG

Retrieval Augmented Generation (RAG) means “stuffing the prompt with relevant information” to improve the model’s response. This is a 2 step process:

1. Retrieve relevant information from the knowledge base.
2. Augment the prompt to provide context to the model.

Let’s build a traditional RAG Agent that answers questions from a PDF of recipes.

1. Install libraries

   Install the required libraries using pip

   Mac

   pip install -U pgvector pypdf "psycopg[binary]" sqlalchemy
   

2. Create a Traditional RAG Agent

   Create a file traditional_rag.py with the following contents

   traditional_rag.py

   from agno.agent import Agent
   from agno.models.openai import OpenAIChat
   from agno.knowledge.pdf_url import PDFUrlKnowledgeBase
   from agno.vectordb.pgvector import PgVector, SearchType
   
   db_url = "postgresql+psycopg://ai:ai@localhost:5532/ai"
   knowledge_base = PDFUrlKnowledgeBase(
       # Read PDF from this URL
       urls=["https://agno-public.s3.amazonaws.com/recipes/ThaiRecipes.pdf"],
       # Store embeddings in the `ai.recipes` table
       vector_db=PgVector(table_name="recipes", db_url=db_url, search_type=SearchType.hybrid),
   )
   # Load the knowledge base: Comment after first run
   knowledge_base.load(upsert=True)
   
   agent = Agent(
       model=OpenAIChat(id="gpt-4o"),
       knowledge=knowledge_base,
       # Enable RAG by adding references from AgentKnowledge to the user prompt.
       add_references=True,
       # Set as False because Agents default to `search_knowledge=True`
       search_knowledge=False,
       markdown=True,
       # debug_mode=True,
   )
   agent.print_response("How do I make chicken and galangal in coconut milk soup")
   

3. Run the agent

   Run the agent (it takes a few seconds to load the knowledge base).

   Mac

   python traditional_rag.py
   

   
   

How to use local PDFs

If you want to use local PDFs, use a PDFKnowledgeBase instead

agent.py

from agno.knowledge.pdf import PDFKnowledgeBase

...
knowledge_base = PDFKnowledgeBase(
    path="data/pdfs",
    vector_db=PgVector(
        table_name="pdf_documents",
        db_url=db_url,
    ),
)
...

Step 3: Agentic RAG

With traditional RAG above, add_references=True always adds information from the knowledge base to the prompt, regardless of whether it is relevant to the question or helpful.

With Agentic RAG, we let the Agent decide if it needs to access the knowledge base and what search parameters it needs to query the knowledge base.

Set search_knowledge=True and read_chat_history=True, giving the Agent tools to search its knowledge and chat history on demand.

1. Create an Agentic RAG Agent

   Create a file agentic_rag.py with the following contents

   agentic_rag.py

   from agno.agent import Agent
   from agno.models.openai import OpenAIChat
   from agno.knowledge.pdf_url import PDFUrlKnowledgeBase
   from agno.vectordb.pgvector import PgVector, SearchType
   
   db_url = "postgresql+psycopg://ai:ai@localhost:5532/ai"
   knowledge_base = PDFUrlKnowledgeBase(
       urls=["https://agno-public.s3.amazonaws.com/recipes/ThaiRecipes.pdf"],
       vector_db=PgVector(table_name="recipes", db_url=db_url, search_type=SearchType.hybrid),
   )
   # Load the knowledge base: Comment out after first run
   knowledge_base.load(upsert=True)
   
   agent = Agent(
       model=OpenAIChat(id="gpt-4o"),
       knowledge=knowledge_base,
       # Add a tool to search the knowledge base which enables agentic RAG.
       search_knowledge=True,
       # Add a tool to read chat history.
       read_chat_history=True,
       show_tool_calls=True,
       markdown=True,
       # debug_mode=True,
   )
   agent.print_response("How do I make chicken and galangal in coconut milk soup", stream=True)
   agent.print_response("What was my last question?", markdown=True)
   

2. Run the agent

   Run the agent

   Mac

   python agentic_rag.py
   

   Notice how it searches the knowledge base and chat history when needed

Attributes

Parameter	Type	Default	Description
knowledge	AgentKnowledge	None	Provides the knowledge base used by the agent.
search_knowledge	bool	True	Adds a tool that allows the Model to search the knowledge base (aka Agentic RAG). Enabled by default when knowledge is provided.
add_references	bool	False	Enable RAG by adding references from AgentKnowledge to the user prompt.
retriever	Callable[..., Optional[list[dict]]]	None	Function to get context to add to the user message. This function is called when add_references is True.
context_format	Literal['json', 'yaml']	json	Specifies the format for RAG, either “json” or “yaml”.
add_context_instructions	bool	False	If True, add instructions for using the context to the system prompt (if knowledge is also provided). For example: add an instruction to prefer information from the knowledge base over its training data.

Developer Resources

• View Cookbook