Citations in ADK

[VenkataSakethDakuri]

Is there any native way of getting citations for the Agent's response ?

[adk-bot]

Response from ADK Answering Agent (experimental, answer may be inaccurate)

TLDR: Yes, ADK provides a native way to get citations for an agent's response, especially when using tools like VertexAiSearchTool that ground the response in a knowledge base. You can achieve this by implementing an after_model_callback to process the grounding_metadata from the LlmResponse.

---

Hello! Yes, it's possible to retrieve citations for an agent's response in ADK. When an agent's response is grounded in data from a tool like Vertex AI Search, the citation information is available in the grounding_metadata of the LlmResponse.

You can define an after_model_callback function for your agent to process this metadata and append the citations to the response. Here's an example of how you can implement this:

def citation_retrieval_after_model_callback(
    response: LlmResponse,
) -> LlmResponse:
  """Adds citations to the response if grounding metadata is present."""
  if response.grounding_metadata and response.grounding_metadata.citations:
    response.text += "\n\nCitations:\n"
    for citation in response.grounding_metadata.citations:
      response.text += f"  - {citation.source_id}: {citation.text}\n"
  return response

# When creating your agent, you can add this callback:
my_agent = LlmAgent(
    ...
    after_model_callbacks=[citation_retrieval_after_model_callback],
)

This approach allows you to automatically include citations, which helps in building trust and providing traceability for the information provided by the agent. You can find this example and more details in the ADK documentation. [1]

[1] https://github.com/google/adk-python/blob/main/contributing/samples/adk_knowledge_agent/agent.py

[VenkataSakethDakuri]

Is it compulsory to use any specific tool like VertexAI or Google Search for the citations ? Or does it work generally ?