guepard-relml

A natural-language agent over your relational database. Ask a question in plain English — it explores your schema, frames a predictive task, trains a model, and answers you.

Install

Requires Python 3.9+. Install into a virtual environment:

python3 -m venv .venv && source .venv/bin/activate
pip install guepard-relml

On Debian/Ubuntu, error: externally-managed-environment just means the venv step was skipped — the two commands above are the fix.

Configure the model backend — AWS Bedrock

The agent runs Claude on AWS Bedrock. Set two environment variables (or put them in a .env file in your working directory):

export RELML_AGENT_BACKEND=bedrock
export AWS_BEARER_TOKEN_BEDROCK=your-bedrock-api-key
# optional:
export AWS_REGION=us-east-1                       # default
export RELML_AGENT_MODEL=us.anthropic.claude-sonnet-4-5-20250929-v1:0

Create the key in the AWS Bedrock console (Bedrock → API keys), and make sure Claude models are enabled for your account and region.

Other backends: the agent can also use the Anthropic API directly (RELML_AGENT_BACKEND=anthropic, ANTHROPIC_API_KEY=…) or Ollama (RELML_AGENT_BACKEND=ollama, OLLAMA_API_KEY=…).

Use it — Python

source can be a folder of CSVs, a CSV/Parquet file, or a Postgres DSN.

from guepard.tools.agent import Agent

agent = Agent("./data")
answer = agent.ask("Which customers are most likely to churn next month?")
print(answer)

Quiet mode (no live progress output) — handy in scripts:

agent = Agent("./data", verbose=False)
print(agent.ask("Forecast next week's daily order volume."))

Choose the backend explicitly in code:

from guepard.tools.agent import Agent, LLMClient

agent = Agent("./data", client=LLMClient(backend="bedrock"))
print(agent.ask("Rank drivers by DNF risk for the next race."))

Use it — command line

relml-agent --source ./data                                        # interactive REPL
relml-agent "who is likely to churn next month?" --source ./data   # one-shot

Inside the REPL, /help lists commands (/tables, /schema, /sql, /models, /predict, /plot, …).

License

MIT