Banko AI Assistant

A demo banking assistant that combines vector RAG, a LangGraph multi-agent pipeline, and runtime-switchable LLM providers — all backed by a single CockroachDB cluster.

What it can do

• Ask in natural language — "what did I spend on dining last month?", "show me anything unusual" — answers come from vector search over your expenses plus an LLM-generated summary.
• Upload a receipt — image or PDF goes through an OCR + agent pipeline (Receipt → Fraud → Budget) that extracts items, flags duplicates, and reports budget impact, with each agent step durably checkpointed.
• Multi-provider LLM — watsonx (default), OpenAI, AWS Bedrock, or Google Gemini. The provider is chosen at startup via AI_SERVICE; within a running app you can swap models inside the active provider from Settings (model lists come from each provider's API, not a hardcoded enum). Switching providers requires a restart with the new env value.
• See agents work in real time — the dashboard streams actual agent state from the backend (no canned activity).
• Persistent chat — conversations survive restarts via CockroachDBChatMessageHistory.
• Three-layer semantic cache — query / embedding / vector-search caches with tunable similarity thresholds.

How it works

Five layers, one database:

All CockroachDB-specific pieces come from langchain-cockroachdb: CockroachDBEngine (psycopg3 pool), CockroachDBVectorStore, CockroachDBChatMessageHistory, and CockroachDBSaver.

Run it

Prerequisites

• Python 3.10+ (3.12 recommended)
• CockroachDB v25.4.0+ (vector indexes are GA)
• An API key for at least one provider (watsonx, OpenAI, AWS, or Gemini), or a local model through Ollama (no key, no internet)

Install

pip install banko-ai-assistant            # PyPI
# or
docker-compose up -d                      # Docker
# or
git clone https://github.com/cockroachlabs-field/banko-ai-assistant
cd banko-ai-assistant
uv pip install -e ".[dev]"                # local development

Start CockroachDB

brew install cockroachdb/tap/cockroach    # macOS
cockroach start-single-node --insecure \
  --store=./cockroach-data \
  --listen-addr=localhost:26257 \
  --http-addr=localhost:8080 --background

Start the app

export AI_SERVICE=watsonx                 # or openai, aws, gemini, ollama
export WATSONX_API_KEY=...
export WATSONX_PROJECT_ID=...
export DATABASE_URL="cockroachdb://root@localhost:26257/defaultdb?sslmode=disable"
export CDC_WEBHOOK_HMAC_SECRET=dev-only-secret   # signs Coach webhook posts, enables the welcome nudge

banko-ai run                              # port 5000, generates 5000 sample records on first start
banko-ai run --port 5001                  # custom port (macOS AirPlay grabs 5000)
banko-ai run --no-data                    # skip the sample-data generator

Open http://localhost:5000.

Sign up and it's yours

The first visit asks for a username. New names pick a spending style (diner, subscriber, saver, balanced) and get a seeded history in that style, embeddings included; returning names land straight back in their own data, and the Coach greets new users with a real nudge within seconds. Everything on screen is scoped to that identity: SQL aggregations filter by your user id, vector search runs through the per-user vector index, and the header shows who is signed in. Questions that ask for totals or counts ("how much did I spend on restaurants in the past 60 days?") are answered by SQL directly, so the figures are exact and identical no matter which AI provider is active; the model adds the narrative and suggestions around them. Every answer wears a badge with the real database time, and an expandable panel shows the actual EXPLAIN ANALYZE that produced it.

On a multi-region cluster the same app goes region aware with no extra configuration: signup offers a home region (detected live from the cluster), rows are pinned to it via REGIONAL BY ROW, reads prune to the user's partition, and the answer badge names the regions that actually served each query. Kill a region and the badge visibly moves to the survivors while answers keep flowing. Single-region deployments see none of this; nothing fake is ever rendered. The legacy personas (maya, sam, riley) still exist for existing databases, and banko-ai clear-demo-users resets a demo machine while keeping them.

banko-ai --help lists the rest (generate-data, clear-data, status, search, etc.). The first run creates the schema (expense, agent, cache, checkpoint tables), generates sample data with embeddings, and initializes the selected provider.

Run it offline (airgap)

The whole stack runs without internet: embeddings are computed locally, and the LLM is a local model served by Ollama. Bring the stack up once while online and it preloads itself (an init container pulls the Ollama model into a named volume, skipping the pull when it is already cached); every start after that works with the network off:

docker compose -f docker-compose.airgap.yml up -d   # first run online, then works offline

To cache ahead of time instead (say, the night before a talk), scripts/airgap/preload-models.sh pulls the same models explicitly.

Default model is granite3.3:8b (override with OLLAMA_MODEL). For a non-Docker setup, ollama serve plus AI_SERVICE=ollama banko-ai run does the same thing. Any OpenAI-compatible endpoint also works via OPENAI_BASE_URL with AI_SERVICE=openai.

Configuration

The important knobs:

Provider keys depend on what you pick:

# IBM watsonx (default)
WATSONX_API_KEY=...   WATSONX_PROJECT_ID=...

# OpenAI
OPENAI_API_KEY=...

# AWS Bedrock — note AI_SERVICE=aws, not bedrock
AWS_ACCESS_KEY_ID=... AWS_SECRET_ACCESS_KEY=... AWS_REGION=us-east-1

# Google Gemini (Vertex AI)
GOOGLE_APPLICATION_CREDENTIALS=path/to/sa.json   GOOGLE_PROJECT_ID=...
# or the Generative AI API
GOOGLE_API_KEY=...

Override model lists with WATSONX_MODELS, OPENAI_MODELS, AWS_MODELS, GEMINI_MODELS (comma-separated). Cache, fraud, and pool tuning live in banko_ai/config/ (CACHE_SIMILARITY_THRESHOLD, CACHE_TTL_HOURS, FRAUD_DUPLICATE_WINDOW_DAYS, DB_POOL_SIZE, …).

API

Full list with examples in docs/API.md. Quick check:

curl -X POST http://localhost:5000/api/rag \
  -H "Content-Type: application/json" \
  -d '{"query": "What are my biggest expenses this month?"}'

The Spending Coach (streaming + agentic)

Everything above waits for a question. The Coach reacts to events: a spending signal arrives, an agent investigates with real budget and transaction tools, and a nudge appears live on the /coach tab, for example "you've used 82% of your dining budget with 9 days left." Signals and nudges live in CockroachDB with row-level TTLs, and the agent's conversation state checkpoints there too.

Kick it off the quick way, by posting a synthetic signal at the webhook:

export CDC_WEBHOOK_HMAC_SECRET=dev-only-secret   # same value as the app
uv run python scripts/coach/mock_signals.py --type=budget_threshold
# also: --type=anomaly, --type=recurring_drift

The coach page shows the signed-in user's nudges. By default the script targets the maya persona, so sign in as maya to watch the nudge land, or pass --user-id (or set COACH_DEFAULT_USER_ID) for the account you signed up with.

Or run the real thing: CockroachDB changefeeds streaming through Debezium and Kafka into the app. One script brings up Kafka, Kafka Connect, and the Debezium CockroachDB connector, then registers a source on the spending_signals table:

scripts/coach/cdc-demo/run-cdc-demo.sh
# then run the app with the Kafka transport on:
COACH_KAFKA_ENABLED=true KAFKA_BOOTSTRAP_SERVERS=localhost:29092 \
  AI_SERVICE=watsonx banko-ai run

With that stack up, sending a change event is just SQL. Insert a row and CockroachDB streams it to the Coach, no webhook involved. The coach page shows the signed-in user's nudges, so target the username you signed up with:

INSERT INTO spending_signals
  (user_id, signal_type, severity, payload, idempotency_key)
VALUES
  ((SELECT user_id FROM users WHERE username = 'YOUR-USERNAME'),
   'budget_threshold', 'warn',
   '{"category": "dining", "pct_used": 0.82, "monthly_budget": 400.0,
     "spent_so_far": 328.0, "days_remaining": 9}',
   'demo:' || gen_random_uuid()::STRING);

Watch /coach while you run it. To verify the whole path automatically:

uv run python scripts/coach/assert_nudges.py             # webhook transport
uv run python scripts/coach/assert_nudges.py --via sql   # the CDC pipeline

The producer contract (payload shapes, idempotency, both transports) is in PIPELINE_CONTRACT.md.

Where the data plane comes from

This repo is the agent side. A companion repo, viragtripathi/cockroachdb-watsonx-data-pipeline, streams CockroachDB CDC events into Apache Iceberg on IBM watsonx.data (both via webhook and Debezium → Kafka). Neither repo requires the other — this app works against a local CockroachDB with its own sample data — but the two together demo the end-to-end transactional + lakehouse path.

Testing

python -m pytest tests/ -v                # all tests
ruff check banko_ai/                      # lint

Integration tests need a populated CockroachDB; they're skipped in CI when the DB isn't available.

Troubleshooting

• CockroachDB version — must be v25.4.0+ for vector indexes (cockroach version).
• DB connection error — confirm the single-node command above is running, then cockroach sql --insecure --execute "SHOW TABLES;".
• AI provider issues — confirm keys are exported, then hit /api/health. For watsonx, /diagnostics/watsonx has connection details.
• Port 5000 in use (macOS) — AirPlay Receiver claims port 5000. Either disable it in System Settings → AirDrop & Handoff, or run banko-ai run --port 5001.

License

MIT