Knowledge Management

KnowledgeManagementTools enable a builder or operator agent to load content into a knowledge base, inspect what is loaded, and remove it. This is the write side of knowledge: end-user-facing agents get search — give them Knowledge Tools — while this toolkit goes to the agent that curates the base. The knowledge base must have a contents_db: every tool here reads or writes content rows, and construction raises a ValueError without one.

Prerequisites

The following example requires the agno, openai, and qdrant-client libraries, and sqlalchemy for the SQLite contents database (agno[sqlite]). It connects to Qdrant on localhost:6333.

uv pip install -U "agno[sqlite]" openai qdrant-client
docker run -d --name qdrant -p 6333:6333 qdrant/qdrant:latest

Set your OpenAI API key:

export OPENAI_API_KEY=***

Example

The following operator agent loads a website into the knowledge base and reports what is loaded:

cookbook/91_tools/knowledge_management_tools.py

import asyncio

from agno.agent import Agent
from agno.db.sqlite import SqliteDb
from agno.knowledge.embedder.openai import OpenAIEmbedder
from agno.knowledge.knowledge import Knowledge
from agno.models.openai import OpenAIResponses
from agno.tools.knowledge import KnowledgeManagementTools
from agno.vectordb.qdrant import Qdrant

knowledge = Knowledge(
    name="Product Docs",
    contents_db=SqliteDb(db_file="tmp/knowledge_contents.db"),
    vector_db=Qdrant(
        collection="product-docs",
        url="http://localhost:6333",
        embedder=OpenAIEmbedder(id="text-embedding-3-small"),
    ),
)

operator = Agent(
    name="Knowledge Operator",
    model=OpenAIResponses(id="gpt-5.6-luna"),
    tools=[KnowledgeManagementTools(knowledge=knowledge, max_pages=25)],
    markdown=True,
)


async def main() -> None:
    await operator.aprint_response("Load https://docs.agno.com into the knowledge base.")
    await operator.aprint_response("What do we have loaded now?")


if __name__ == "__main__":
    asyncio.run(main())

Content Ownership

scope decides who can read what the operator loads. With scope="shared" (the default), ingested rows are readable by every agent on the knowledge base. With scope="user", rows belong to the run’s user_id — and a run without a user_id gets an error back instead of silently writing to the shared bucket, where the content would be readable and deletable by everyone.

Toolkit Params

Parameter	Type	Default	Description
`knowledge`	`Knowledge`	-	Knowledge base to manage (required). Must have a `contents_db`; construction raises a `ValueError` without one.
`scope`	`Literal["shared", "user"]`	`"shared"`	Ownership of ingested rows. See Content Ownership.
`max_pages`	`int`	`50`	Default page cap per site ingest. Clamped between 1 and 500.
`page_fetcher`	`Optional[Any]`	`None`	Custom page fetcher. The default resolves the Parallel fetcher when available, and the built-in fetcher otherwise.
`ingest_url`	`bool`	`True`	Register the `ingest_url` tool.
`ingest_path`	`bool`	`False`	Register the `ingest_path` tool. Off by default — see the warning below.
`ingest_text`	`bool`	`True`	Register the `ingest_text` tool.
`remove_content`	`bool`	`True`	Register the `remove_content` tool, gated behind confirmation by default.
`instructions`	`Optional[str]`	`None`	Custom instructions for the operator agent. When unset, the toolkit’s built-in operator instructions are used.
`add_instructions`	`bool`	`True`	Whether to add the instructions to the agent’s context.

list_content and ingest_status are always registered — they only read.

ingest_path is off by default, and registering it is an explicit choice: it reads any path the server process can read, and under scope="shared" everything it loads becomes readable by every agent on that knowledge base. Enable it only where the operator agent is trusted with the machine’s filesystem.

Toolkit Functions

Function	Description
`ingest_url`	Ingest a website from its sitemap, one row per page, up to `max_pages` (an optional per-call `max_pages` overrides the default). Re-running refreshes changed pages, retries failed ones, and prunes pages that left the sitemap.
`ingest_path`	Ingest a file or folder from the server’s filesystem. A folder lands one row per file, refreshed by content digest. Registered only when `ingest_path=True`.
`ingest_text`	Ingest a named block of text, with optional metadata.
`list_content`	List everything on the knowledge base, grouped by site, with an optional `host` filter.
`ingest_status`	Report a site’s ingestion status, including failed pages.
`remove_content`	Remove a content row by id; removing a site row also removes every page under it. Requires user confirmation by default, and a caller-supplied `requires_confirmation_tools` list is merged with it rather than replacing it.

Every tool returns JSON with an ok field; failures come back as {"ok": false, "error": ...} instead of raising to the model. Each tool also has an async variant registered under the same name, so the async versions are used automatically with arun and aprint_response.

Get Started

Basics

Advanced

Production

Providers

Other

Additional Resources

Knowledge Management

Prerequisites

Example

Content Ownership

Toolkit Params

Toolkit Functions

Developer Resources

​Prerequisites

​Example

​Content Ownership

​Toolkit Params

​Toolkit Functions

​Developer Resources

Prerequisites

Example

Content Ownership

Toolkit Params

Toolkit Functions

Developer Resources