Skip to main content

Quickstart

1. Install​

pip install driftguard-ai
python -m spacy download en_core_web_sm

2. Record a mistake​

from driftguard import DriftGuard

guard = DriftGuard()

guard.record(
    action="increase salt",
    feedback="too salty",
    outcome="dish ruined",
)

3. Review before acting​

review = guard.before_step("add more salt")

if review.warnings:
    print(review.warnings[0].risk)
    # → "too salty"
    print(review.confidence)
    # → 0.82

DriftGuard matched "add more salt" semantically to "increase salt" and surfaced the stored warning.

4. Check graph stats​

print(guard.stats())
# → {'nodes': 3, 'edges': 2}

5. Prune stale memories​

guard.prune()

6. Run the MCP server​

driftguard-mcp

What just happened​

  • record() stored the causal chain action → feedback → outcome into the in-memory semantic graph and persisted it to disk.
  • before_step() embedded the query, retrieved similar action nodes, walked their causal chains, and returned ranked warnings.
  • The graph persists across restarts — DriftGuard remembers between sessions.

Next steps​