quickstart_csv — CSV cold start

The most common FDE first-day scenario: the customer hands you a CSV export and nothing else. This example walks the full 5-step FDE workflow against a tiny 20-row sample.

Run it

# 1. install (core layer only — no agentscope, no docker, no API key)
pip install -e ".[dev]"

# 2. connect — preview the schema + sample
fde-scope connect --type csv --source examples/quickstart_csv/sample_tickets.csv

# 3. forge — turn raw rows into an auditable corpus report
fde-scope corpus \
  --input examples/quickstart_csv/sample_tickets.csv \
  --config fde_scope/templates/corpus_config.yaml \
  --out reports/corpus_report.html

# 4. deploy — assemble a tenant agent (dry-run; no docker needed)
fde-scope deploy --tenant client_a --dry-run

# 5. eval — benchmark a (mock) agent over an eval set
fde-scope eval --agent mock --test-set <path-to-eval.jsonl>

# 6. flywheel — replay sample concept events into the data flywheel
fde-scope flywheel --agent client_a --events <path-to-events.json>

Open reports/corpus_report.html to see the auditable deliverable: real vs synthetic counts, coverage gaps, PII masked, and the train/eval/test split.

What this proves

• The core data layer (connect / corpus / eval) runs with zero external dependencies — no LLM, no Docker, no AgentScope.
• The Corpus Engine is genuinely functional in rule-based v0: PII scrubbing, dedup, quality gate, coverage-gap detection, and targeted synthesis all produce real output.
• LLM enhancement (MiMo Token Plan) is strictly optional and layered on top via environment variables — see ../docs/llm_integration.md.
• The runtime layer (deploy / flywheel) is assembled against the real AgentScope 2.0 API — see docs/agentscope_api_mapping.md.