Skip to content

README.md

Latest commit

 

History

History
77 lines (60 loc) · 3.51 KB

README.md

File metadata and controls

77 lines (60 loc) · 3.51 KB
Raw

Examples

Runnable demos showing how to use adk-aerospike.

Script What it shows Needs LLM API key
quickstart.py Minimal LlmAgent wired with AerospikeSessionService and Runner.run_async ✅ Yes — Gemini
with_memory.py Session + lexical memory service end-to-end ✅ Yes — Gemini
adk_samples_wiring.py Smoke-test that Google's own adk-samples work with our three services. Imports each sample's root_agent, constructs Runner(...), drives create_session → append_event → get_session → add_session_to_memory → search_memory. ❌ No
adk_samples_e2e.py Same samples, but actually runs one user turn per agent via Runner.run_async() against a real Gemini call. Dumps the resulting Aerospike state. ✅ Yes — Gemini

Setup for the sample harnesses

# 1. Clone Google's ADK samples next to this repo
git clone --depth 1 https://github.com/google/adk-samples ../adk-samples

# 2. Some samples need extra deps
pip install 'a2a-sdk[all]' mcp

# 3. For E2E only — add your key to .env at the repo root (.env is gitignored)
echo "GEMINI_API_KEY=AIza..." > ../.env

Run

# Wiring-only (no LLM cost, ~30s):
python examples/adk_samples_wiring.py

# End-to-end with real Gemini turns (~2 min, small Gemini cost):
SLACK_MCP_XOXP_TOKEN=dummy python examples/adk_samples_e2e.py
# (the dummy SLACK token avoids an import-time MCPToolset error in one sample)

# Both scripts accept --uri (any Aerospike cluster) and --samples-path.

Compat shims (built-in)

Several official samples have import-time hooks that fail in a clean environment. The harnesses apply two narrowly-scoped shims so the samples can be evaluated against our services:

  1. google.auth.default stubbed during sample import only. Vertex-coupled samples (memory-bank, customer-service, blog-writer) construct Vertex clients at module load and call google.auth.default(), raising DefaultCredentialsError on hosts without GCP Application Default Credentials. We stub the call for the duration of importlib.import_module(...) and restore the real function before any runtime LLM call so we don't poison the auth path.
  2. google.adk.a2a.utils.agent_to_a2a.to_a2a stubbed to a no-op. currency-agent pins a2a-sdk==0.3.3 whose a2a.server.apps module path moved in 1.0+. The unused a2a_app export still constructs.
  3. GOOGLE_GENAI_USE_VERTEXAI=false is set so the SDK uses the Gemini Developer API (GOOGLE_API_KEY path) rather than auto-picking Vertex when ADC happens to be discoverable.

These shims only widen what's importable. They don't paper over runtime problems: a sample that genuinely needs Vertex at runtime would still need real ADC.

Reference outcomes (alpha)

Curated set of 7 samples covering LlmAgent, SequentialAgent, ParallelAgent, multi-agent workflows, and agents with tools, validated against the AWS Pegasus cluster:

  • Wiring: 7 / 7 ✓ — all samples import, accept our service trio via Runner(...), and round-trip a synthetic event through create_session → append_event → get_session → add_session_to_memory → search_memory.
  • E2E: 7 / 7 ✓ — all 7 drive real Gemini turns through runner.run_async(). Cluster state after one run: adk_sessions: 7, adk_memory: 21 (one session per sample, multiple text-bearing events from multi-agent samples like parallel_task which emits 17 events and blog-writer with transfer_to_agent calls).