[RECIPE PROPOSAL] Multi-agent GraphRAG over a Neo4j knowledge graph (Python)

⚠️ Scope Notice

For contrib/ recipes only. core/ recipes follow a different process.

1- What is the intent/purpose of your new recipe? In what way will it help the ADK community? A multi-agent GraphRAG (graph retrieval-augmented generation) recipe: a root orchestrator routes questions to three specialist sub-agents over a Neo4j knowledge graph: (a) a Text-to-Cypher agent that reads the live schema, generates read-only Cypher, and self-corrects; (b) a hand-written graph-tool agent (company → investor lookup); and (c) an agent that calls pre-validated queries via the MCP Toolbox.

It fills a real gap: there is no knowledge-graph / GraphRAG recipe in the repo today, the existing retrieval recipes (rag-agent-search, rag-vector-search, multiformat-hybrid-rag) are all vector-based. It teaches graph-grounded retrieval (multi-hop / relationship queries with explainable results) as a complement to vector RAG, and shows three distinct graph-retrieval strategies in one place. It runs out-of-the-box against Neo4j's public read-only demo database, so it needs only a Gemini API key.

2- What language is your recipe is implemented in (e.g. Python, Go, etc)? Python (3.11+), using google-adk.

3- Will this be a deployable recipe, either by having a Dockerfile, or by utilizing Vertex AI? Not deployable as-is (no Dockerfile). It runs locally via adk web / adk run. It supports Vertex AI as an auth option (GOOGLE_GENAI_USE_VERTEXAI=1) in addition to the Google AI Studio API key. Happy to add a Dockerfile if the team would like it to be deployable.

4- What GCP services or third-party APIs (Salesforce, Jira, etc.) are required?

• Gemini via Google AI Studio API key, or Vertex AI.
• Neo4j, defaults to the public, read-only "companies" demo database (neo4j+s://demo.neo4jlabs.com), so no provisioning is needed to try it.
• Optional: MCP Toolbox (genai-toolbox) for the pre-validated-query agent; the recipe falls back gracefully to schema + Cypher tools when it isn't run.

5- Who (individual/team) will own and maintain this? JOFLOX (individual contributor), listed as ownership.poc in the recipe manifest.

6- Please include any additional information you'd like to share. Already implemented and CI-clean in PR #2616 (validate, ruff, check_env_vars, check_recipe_pyproject, and pytest all pass locally; every automated-review finding across the rounds has been addressed). Adapted from the Apache-2.0 "Building GraphRAG Agents with ADK and Neo4j" codelab, with attribution in the recipe README. Opening this proposal per @happyhuman 's request so the team can approve the recipe before PR review.