Learning & Scoring

flyto-blueprint learns from executed workflows and automatically scores blueprints based on outcomes.

Learning Pipeline

1. Receive executed workflow
2. Compute structural fingerprint (module + param keys)
3. Dedup check → boost existing or create new
4. Abstract concrete values → template arguments
5. Detect compose opportunities (e.g., browser.launch + browser.goto → browser_init)
6. Set initial score (70 if verified, 50 otherwise)
7. Persist to storage

Learning from a Workflow

workflow = {
    "name": "My Pattern",
    "steps": [
        {"id": "s1", "module": "browser.launch", "params": {"headless": False}},
        {"id": "s2", "module": "browser.goto", "params": {"url": "https://example.com"}},
        {"id": "s3", "module": "browser.click", "params": {"selector": "#btn"}},
    ]
}

result = engine.learn_from_workflow(workflow, name="Click Pattern")
# {'ok': True, 'data': {'id': 'click_pattern', 'score': 50, ...}}

Learning from Execution (Verified)

result = engine.learn_from_execution(workflow, name="Verified Pattern")
# Score starts at 70 instead of 50

Scoring System

Event	Score Change
Success reported	+5 (capped at 100)
Failure reported	-10 (floored at 0)
Dedup boost	+3 (same structure learned again)
Auto-retire	When score drops below 10

Reporting Outcomes

engine.report_outcome("click_pattern", success=True)   # Score: 50 → 55
engine.report_outcome("click_pattern", success=True)   # Score: 55 → 60
engine.report_outcome("click_pattern", success=False)  # Score: 60 → 50

Recent reports are deduplicated (1-hour window, max 200 entries).

Deduplication

Two workflows produce the same fingerprint if they have:

• Same module sequence
• Same parameter keys (values are ignored)

Workflow A: browser.goto(url="x") → browser.click(selector="y")
Workflow B: browser.goto(url="z") → browser.click(selector="w")
→ Same fingerprint → existing blueprint boosted by +3

Search & Ranking

Blueprints are ranked by:

Factor	Weight
Tag match	+3 per matching tag
Name match	+2
Description match	+1
Quality score	score / 100

Retired blueprints (score < 10) are excluded from results.

Auto-Retirement

When a blueprint's score drops below 10, it is marked retired: true and no longer appears in search results or listings. This ensures low-quality patterns are automatically cleaned up.