Are citations triggered everytime ?

[VenkataSakethDakuri]

chunks = llm_response.grounding_metadata.grounding_chunks
Do we get these chunks every time the agent generates content, or do they come only when the citations are actually needed to generate the answer instead of giving answer from its internal memory (training) ?

[adk-bot]

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

TLDR: The grounding_chunks are provided only when the model's response is based on information retrieved from external data sources. They are not returned when the model answers from its internal memory.

---

Hello! That's an excellent question.

The grounding_metadata attribute, which contains the grounding_chunks, is populated only when the model determines that it needs to consult an external data source to provide a high-quality answer. This happens when the query requires up-to-date information or specific knowledge that isn't part of the model's internal training data. The model then uses a grounding tool, such as Vertex AI Search or Google Search, to fetch this information. [1, 2]

If the model can answer the question using its own internal knowledge, it will not invoke a grounding tool, and therefore the llm_response object will not contain the grounding_metadata attribute. In the ADK, the grounding_metadata field is defined as an optional attribute on the LlmResponse object, meaning it can be None. [3]

It is a good practice to check if grounding_metadata exists before attempting to access it, as shown in the code documentation. [4]

[1] Grounding with Vertex AI Search
[2] Grounding with Google Search
[3] llm_response.py on GitHub
[4] agent.py on GitHub