⚠️ IMPORTANT: All features are experimental, under active development. Use at your own risk. Customization to your workflow required. © 2026 GLG, a.s. | ← Back to Index

4. Focus Engine — Smart Context Selection

The Focus Engine is UAML's key differentiator. Instead of dumping all memories into context, it intelligently selects the most relevant ones within a token budget.

4.1 How It Works

Your Query → Focus Engine → Ranked, filtered, budget-constrained context
                 ↑
         Configuration (presets or custom rules)

4.2 Input Filter (What Gets Stored)

6 sequential filter stages protect data quality:

Stage	Filter	What It Does	Configurable?
1	Length filter	Rejects entries < N chars	min_entry_length (default: 10)
2	Token limit	Rejects entries > N tokens	max_entry_tokens (default: 2000)
3	PII detector	Detects emails, phones, IDs, credit cards	Per-type: allow/mask/reject
4	Category filter	Per-category policies	allow/deny/encrypt/require_consent
5	Rate limiter	Token-bucket rate limiting	rate_limit_per_min (default: 100)
6	Relevance gate	Rejects low-quality entries	min_relevance_score (default: 0.3)

4.3 Output Filter (What Gets Recalled)

The Focus Engine scores and ranks memories using:

• Relevance to the query (semantic similarity)
• Freshness (exponential time decay)
• Confidence (source reliability)
• Token budget (never exceed the context window allocation)

4.4 Presets

Three built-in presets — choose based on your use case:

Preset	Description	Best For
Conservative	Strict filtering, high relevance threshold	Production, customer-facing
Standard	Balanced filtering (default)	General use
Research	Loose filtering, more context	Exploration, analysis

# Apply a preset
uaml.apply_preset("conservative")

# Or via REST
curl -X PUT http://localhost:8780/api/v1/focus-config \
  -d '{"preset": "standard"}'

4.5 Custom Rules (Starter+)

Fine-tune individual parameters:

config = {
    "min_relevance_score": 0.5,    # 0.0–1.0 (higher = stricter)
    "max_entry_tokens": 1500,       # max tokens per entry
    "freshness_decay": 0.95,        # time decay factor
    "min_confidence": 0.6,          # minimum confidence threshold
    "pii_policy": "mask",           # allow | mask | reject
    "categories": {
        "personal": "encrypt",
        "work": "allow",
        "sensitive": "require_consent"
    }
}

4.6 Rules Changelog (Pro+)

Every configuration change is logged:

GET /api/v1/rules-log?limit=20

Returns: who changed what, when, expected impact. After 7 days, automatic evaluation compares predicted vs. actual impact.

---