Rasa Maestro Voice Banking Skills

Build a voice retail banking agent with the new Rasa Skills / Maestro architecture and Deepgram for speech-to-text and text-to-speech.

We call our agent Rasano.

This repo is both:

1. A finished working agent you can run immediately
2. A live-session tutorial (tutorial/TUTORIAL.md) with paste-ready snippets

It replaces the older CALM-flows retail banking starter with Skills, progressive control, and end-to-end voice.

What Rasano can do

Skill	Capability
check_balance	Look up account balances
transfer_money	Send money (composes add_payee when needed)
list_payees / add_payee / remove_payee	Manage authorised payees
block_card	Block lost/stolen/damaged cards + replacement
banking_faq	Answer common banking questions from references
human_handoff	Create a live-agent ticket
goodbye / intro	Orient and close the conversation

Demo customer: John Smith (seeded SQLite bank under data/source/).

Prerequisites

• Python 3.11 or 3.12
• uv
• Rasa Pro Developer Edition license (RASA_LICENSE)
• OpenAI API key (OPENAI_API_KEY)
• Deepgram API key (DEEPGRAM_API_KEY) for ASR and TTS

Quick start

make install     # install dependencies with uv
make env         # create .env from .env.example
# edit .env and set RASA_LICENSE, OPENAI_API_KEY, DEEPGRAM_API_KEY

make verify      # pre-flight: keys, project, demo data, connectivity
make train
make inspect

make verify is the single command to run whenever anything looks wrong. It checks your Python version, license expiry, API keys, project validity, demo data, and live connectivity to OpenAI and Deepgram, and tells you the exact fix for anything it finds.

make inspect opens the Inspector. Use the microphone for voice (Deepgram) or type as a fallback.

Project layout

agent.yml              # Rasano persona + voice flags
integrations.yml       # OpenAI + Inspector (Deepgram ASR/TTS)
memory.yml             # project-wide memory
responses.yml          # greeting override
skills/                # one folder per skill (skill-local tools.py lives here)
tools/                 # shared @tool functions (used by 2+ skills)
lib/                   # SQLite demo bank helpers
data/source/           # seed JSON
tutorial/              # live session script + snippets

Live tutorial

• Audience guide: tutorial/TUTORIAL.md
• Paste sets: tutorial/snippets/

Teaching spine: scaffold → FAQ prose → first tool → tool constraints → block-card progressive control → transfer composition → voice demo.

Progressive control (quick map)

Lever	Where	What it guarantees
tool_constraints.requires	skill frontmatter	Tool hidden until memory condition is true
requires_confirmation	skill frontmatter	Explicit user approval before side effects
if: paragraphs	skill body	Only matching branch stays in the LLM prompt
utter: + responses.yml	frontmatter + YAML	Exact wording (compliance / warnings)
:::ordered_block	skill body	Strict step order when sequence is the requirement

Makefile

Run make with no arguments for the full grouped help screen.

Target	Action
make install	Install dependencies (uv sync --prerelease=allow)
make env	Create .env from .env.example (never overwrites)
make verify	Full pre-flight diagnostics — start here if stuck
make validate	Fast skill / memory / tool validation only
make train	Validate + package the agent
make inspect	Open Inspector (voice + text)
make run	Start the API server
make show-demo-data	Print John Smith's accounts, cards, payees
make reset-db	Reseed the demo bank from data/source/
make tutorial	Show the live-session chapters and snippet paths
make clean	Remove models / caches / local db
make clean-all	Also remove .venv (full reset)

Notes for Rasa 3.19.0.dev3

• Package: rasa-pro==3.19.0.dev3 (pre-release; uv prereleases enabled)
• Scaffold a fresh Maestro project with rasa init --engine maestro (there is no --template voice)
• LLM is gpt-5.2; do not set temperature (GPT-5 reasoning models only accept the default)
• Tools import from rasa.calm_v2.tools (not classic rasa_sdk Action classes)
• Local-first tools: a single-skill tool lives in skills/<name>/tools.py and is auto-discovered; only tools shared by 2+ skills go in tools/*.py and are pulled in with import_tools
• Tools are named in plain prose (Call list_accounts), never @tool.name
• Identity is loaded deterministically at session start: the project overrides default_session_start to execute_tool-run load_customer_profile
• Channels live in integrations.yml (no credentials.yml for Maestro projects)
• Inspector uses Deepgram by default when configured under channels.inspector

License

Use with a valid Rasa Pro Developer Edition license. Review the Developer Terms.