Knowledge

Build knowledge bases with vector storage, embeddings, and content ingestion for retrieval-augmented generation (RAG).

Content (text, text/Markdown/JSON URLs)
  └─▶ Knowledge
        └─▶ VectorDB (Qdrant)
              └─▶ Embedder (OpenAI)

​

agno/knowledge

​

Config

​

Outputs

​

agno/vectordb/qdrant

​

Config

​

Outputs

​

agno/knowledge/embedder/openai

​

Config

​

Outputs

​

agno/knowledge/content

​

Config

​

Outputs

​

Example

# 1. Embedder
provider: agno
resource: knowledge/embedder/openai
name: embedder
config:
  id: text-embedding-3-small
  api_key:
    provider: pragma
    resource: secret
    name: openai-key
    field: outputs.OPENAI_API_KEY
---
# 2. Vector database
provider: agno
resource: vectordb/qdrant
name: vectors
config:
  url:
    provider: qdrant
    resource: database
    name: main
    field: outputs.url
  collection:
    provider: qdrant
    resource: collection
    name: docs
    field: outputs.name
  api_key:
    provider: qdrant
    resource: database
    name: main
    field: outputs.api_key
  search_type: vector
  embedder:
    provider: agno
    resource: knowledge/embedder/openai
    name: embedder
---
# 3. Knowledge base
provider: agno
resource: knowledge
name: docs-kb
config:
  vector_db:
    provider: agno
    resource: vectordb/qdrant
    name: vectors
  max_results: 5
---
# 4. Content source
provider: agno
resource: knowledge/content
name: user-guide
config:
  knowledge:
    provider: agno
    resource: knowledge
    name: docs-kb
  url: "https://docs.example.com/guide.md"
  topics:
    - documentation
    - user-guide

​

Notes

• You must provide exactly one of url or text_content in content resources.
• Content reads inline text, or a URL of a plain text, Markdown or JSON file whose path ends in .txt, .text, .md, .markdown or .json. Web pages, PDF, Office, CSV and YouTube URLs are refused, and a URL that serves an HTML page fails. A JSON file must hold an object or array at its top level.
• search_type accepts only vector: a qdrant/collection holds one unnamed dense vector, and keyword or hybrid search needs sparse vectors.
• The Qdrant collection must exist before content is written. Apply a qdrant/collection sized for the embedder’s vector dimensions first; a content resource never creates the collection.
• Content resources are stateful — create and update write the content’s chunks into the collection and remove any older chunks the resource wrote; delete removes every chunk the resource wrote.
• The knowledge reference on content, the vector_db reference on a knowledge base, and the url and collection of a vector database are immutable — delete and recreate the resource to change them.