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())
# → {'mistakes': {'nodes': 3, 'edges': 2}, 'successes': {'nodes': 0, 'edges': 0}}

guard.stats() reports the mistake graph and the success graph separately — see Success Memory for more on the latter.

5. Record a success​

guard.record_success(
    action="add more salt",
    feedback="well seasoned",
    outcome="dish praised",
)

review = guard.before_step("add a pinch of salt")

for reinforcement in review.reinforcements:
    print(reinforcement.trigger, "->", reinforcement.recommendation)
    # → "add more salt" -> "well seasoned"

6. Prune stale memories​

guard.prune()

7. Run the MCP server​

driftguard-mcp

What just happened​

  • record() stored the causal chain action → feedback → outcome into the in-memory mistake graph and persisted it to disk.
  • before_step() embedded the query, retrieved similar action nodes from both graphs, walked their causal chains, and returned ranked warnings (from the mistake graph) and reinforcements (from the success graph).
  • record_success() stored a positive causal chain into the separate success graph.
  • Both graphs persist across restarts — DriftGuard remembers between sessions.

Next steps​