Create VertexRagRetrievalTool

[federicobassetto]

Let's add the VertexRagRetrievalTool like in the python sdk https://google.github.io/adk-docs/integrations/vertex-ai-rag-engine/

[Varun-S10]

Hi @federicobassetto, Thanks for bringing this up! I will escalate the request to add the VertexRagRetrievalTool to adk-js to the team.

Hi @tpryan, user is requesting the addition of a VertexRagRetrievalTool to the adk-js. This feature would enable private data retrieval using Vertex AI RAG Engine, mirroring the functionality already available in the Python ADK. The Python SDK currently offers a VertexRagRetrievalTool (as documented here: link) .

[federicobassetto]

Vertex AI API limitation: single corpus per ragResources

While building a custom VertexRagRetrievalTool ahead of an official one, I ran into an important constraint worth flagging for the implementation.

The Vertex AI GenerateContent API only allows one corpus inside retrieval.vertexRagStore.ragResources. Sending multiple VertexRagStoreRagResource entries in a single retrieval tool returns:

tools[0].retrieval.vertex_rag_store.rag_resources: [FIELD_INVALID] only one corpus can be specified

This was the request payload that triggered the error:

{
  "retrieval": {
    "vertexRagStore": {
      "ragResources": [
        { "ragCorpus": "projects/.../ragCorpora/6478428063972458496" },
        { "ragCorpus": "projects/.../ragCorpora/3019663550151917568" }
      ],
      "similarityTopK": 10,
      "vectorDistanceThreshold": 0.6
    }
  }
}

Implication for the official tool design: if a user needs to search across multiple corpora (e.g. a meeting corpus and a document corpus), the tool must either:

1. Create separate retrieval tool entries in llmRequest.config.tools — one per corpus, or
2. Accept only a single corpus and require users to instantiate multiple tool instances (like the Python SDK pattern).

The ragResources field in the @google/genai TypeScript types (VertexRagStore.ragResources?: VertexRagStoreRagResource[]) suggests an array is valid, but the backend rejects it. This mismatch between the SDK types and the API behavior is worth documenting or validating in the official implementation.

[Varun-S10]

Hi @cornellgit, Could you please review this feature request and provide your feedback?

[ScottMansfield]

Taking a look at this

[federicobassetto]

any news?