Customer Support Agent (genai-toolbox)

An LLM-powered customer support service wired to a genai-toolbox toolset (Postgres + Redis) with FastAPI. The agent routes user queries to database lookups, cached answers, or LLM generation, and the toolbox server exposes the underlying tools.

What's included

• app/agent.py: query routing, caching, fallback responses, OpenAI integration.
• app/mcp_client.py: talks to the toolbox server (toolbox_core) and falls back to in-memory mocks.
• app/main.py: FastAPI app exposing POST /support and /health.
• tools.yaml: toolbox configuration for Postgres/Redis tools and the support-toolset.
• TOOLBOX.md: toolbox-specific run instructions.
• OpenAI Agents SDK (openai-agents) drives LLM responses; set OPENAI_API_KEY (and optionally OPENAI_BASE_URL/AGENT_MODEL) to control the backing model.

Prereqs

• Python 3.9+
• Redis running locally (defaults to 127.0.0.1:6379).
• Postgres running locally with a database support_db and user postgres/password (configurable via env).
• Tooling binary: ~/toolbox (genai-toolbox) already downloaded.

Install deps

# (recommended) create/use the project venv
uv init  # if you haven't already; creates .venv and pyproject for uv

# install dependencies into .venv
uv pip install -r requirements.txt

Quick start

1. Copy .env.example to .env, then fill in the values for database, Redis, and your model provider. To target Google Gemini, use:

   OPENAI_API_KEY="$GOOGLE_API_KEY"
   OPENAI_BASE_URL="https://generativelanguage.googleapis.com/v1beta/openai/"
   AGENT_MODEL="gemini-2.5-flash"
   TAVILY_API_KEY=your_tavily_key
    MEMORY_REDIS_URL="redis://127.0.0.1:6379"

2. Activate the virtualenv and export env vars (run this in every shell before starting services):

   cd ~/Documents/projects/customer-support-agent
   source .venv/bin/activate
   set -a && source .env && set +a

3. Start dependencies:

   # Postgres (brew example)
   brew services start postgresql@15
   createdb -h 127.0.0.1 -p 5432 -U postgres support_db
   
   # Redis
   redis-server --port 6379

4. Seed Postgres + Redis with sample support data:

   python scripts/seed_data.py

5. Launch toolbox in its own terminal:

   ~/toolbox --tools-file tools.yaml --port 5000

6. Launch the FastAPI app in another terminal:

   source .venv/bin/activate
   set -a && source .env && set +a
   TOOLBOX_BASE_URL=http://127.0.0.1:5000 uvicorn app.main:app --reload --port 8000

7. Interact with the agent:

   curl -X POST http://127.0.0.1:8000/support \
     -H "Content-Type: application/json" \
     -d '{"query": "What are my recent orders?", "customer_id": "1", "session_id": "cust-1-session"}'

   or

   python tests/test_agent.py

Redis sample data (loaded)

The Redis instance is pre-seeded with support-oriented keys after running python scripts/seed_data.py:

• Customers: support:customer:1..3
• Orders: support:order:1001, support:order:1002, support:order:2001, support:order:3001
• Order lists per customer: support:orders:customer:<id>
• Cached FAQ/last response: support:cache:*

Useful checks:

redis-cli KEYS 'support:*'
redis-cli HGETALL support:customer:1
redis-cli LRANGE support:orders:customer:1 0 -1
redis-cli LRANGE support:memory:cust-1-session 0 -1  # conversation memory

Conversation memory

• Each /support call can include an optional session_id. When provided (or inferred from customer_id), the agent stores the most recent turns in Redis (support:memory:<session>). Those turns are added to the next prompt so the LLM has conversational context.
• Clear session memory with redis-cli DEL support:memory:<session>.
• Gradio (python main.py) exposes separate Customer ID and Session ID fields so you can experiment with different sessions without restarting services.

Troubleshooting

• toolbox failed to initialize: environment variable not found – run set -a && source .env && set +a in every terminal before launching toolbox/uvicorn so ${DB_*}/${REDIS_*} placeholders resolve.
• unable to connect ... user=postgres database=support_db – start Postgres (brew services start postgresql@15 or pg_ctl ... start) and ensure support_db exists. Verify with pg_isready or psql postgres://postgres:password@127.0.0.1:5432/support_db.
• unable to connect to redis ... connect: connection refused – start Redis (redis-server --port 6379 or brew services start redis) and confirm REDIS_ADDRESS/REDIS_URL point at the running instance.
• Agent always replies “I'll help you with that...” – the LLM fallback cached because no API key was available or the provider rejected it. Clear Redis (redis-cli KEYS 'support:*' | xargs redis-cli DEL ...) and ensure OPENAI_API_KEY/OPENAI_BASE_URL are exported before restarting.
• Error getting response: 404 or invalid_api_key – when using Gemini, the key must be OpenAI-compatible. The included OpenAIProvider handles this if you follow the .env instructions above; with a real OpenAI key, leave OPENAI_BASE_URL unset.