Skip to content

~95% on SimpleQA (e.g. Qwen3.6-27B on a 3090). Supports all local and cloud LLMs (llama.cpp, Ollama, Google, ...). 10+ search engines - arXiv, PubMed, your private documents. Everything Local & Encrypted.

2.7 KB
Raw Blame
"""DB-backed context warning checks.

These functions take an open SQLAlchemy session — no Flask dependency.
"""

from typing import Optional

from sqlalchemy import desc, func
from sqlalchemy.orm import Session

from ...database.models import TokenUsage


def check_context_below_history(
    db_session: Session, local_context: int
) -> Optional[dict]:
    """Warn if the current context setting is below what 99% of past researches used.

    Requires at least 5 historical records to produce a meaningful percentile.
    """
    recent_contexts = (
        db_session.query(TokenUsage.context_limit)
        .filter(TokenUsage.context_limit.isnot(None))
        .order_by(desc(TokenUsage.timestamp))
        .limit(100)
        .all()
    )

    if not recent_contexts or len(recent_contexts) < 5:
        return None

    context_values = sorted([r[0] for r in recent_contexts if r[0]])
    if not context_values:
        return None

    # 1st percentile — 99% of researches used at least this much
    percentile_idx = max(0, int(len(context_values) * 0.01))
    min_safe_context = context_values[percentile_idx]

    if local_context >= min_safe_context:
        return None

    return {
        "type": "context_below_history",
        "icon": "📉",
        "title": "Context Below Historical Usage",
        "message": (
            f"Current context ({local_context:,}) is below the context window "
            f"size that 99% of your past researches ran with "
            f"(min safe: {min_safe_context:,}). This may cause truncation."
        ),
        "dismissKey": "app.warnings.dismiss_context_below_history",
        "actionUrl": "/metrics/context-overflow",
        "actionLabel": "View context metrics",
    }


def check_context_truncation_history(
    db_session: Session, local_context: int
) -> Optional[dict]:
    """Warn if past researches experienced truncation at the same or higher context."""
    truncation_count = (
        db_session.query(func.count(TokenUsage.id))
        .filter(TokenUsage.context_truncated.is_(True))
        .filter(TokenUsage.context_limit >= local_context)
        .scalar()
    )

    if not truncation_count or truncation_count <= 0:
        return None

    return {
        "type": "context_truncation_history",
        "icon": "⚠️",
        "title": "Previous Truncation Detected",
        "message": (
            f"Research was truncated {truncation_count} time(s) with similar "
            f"or higher context. Consider increasing context window."
        ),
        "dismissKey": "app.warnings.dismiss_context_truncation_history",
        "actionUrl": "/metrics/context-overflow",
        "actionLabel": "View context metrics",
    }