src/metrics.py

"""
metrics.py — Rolling in-memory metrics for the dashboard panel.

Stores the last 50 query metrics in a deque.
In production, push these to InfluxDB, Prometheus, or BigQuery.
"""

import logging
from collections import deque
from src.utils import safe_divide

logger = logging.getLogger("enterprise-rag.metrics")

_history = deque(maxlen=50)


def record_query_metrics(
    retrieval_latency_ms: float,
    generation_latency_ms: float,
    prompt_tokens: int,
    response_tokens: int,
    eval_scores: dict,
    fallback_used: bool,
):
    """Append one query's metrics to the rolling history."""
    _history.append({
        "retrieval_ms": retrieval_latency_ms,
        "generation_ms": generation_latency_ms,
        "total_ms": retrieval_latency_ms + generation_latency_ms,
        "prompt_tokens": prompt_tokens,
        "response_tokens": response_tokens,
        "total_tokens": prompt_tokens + response_tokens,
        "eval_overall": eval_scores.get("overall", 0),
        "fallback": fallback_used,
    })


def get_metrics_summary() -> str:
    """Build formatted metrics text for the Gradio panel."""
    if not _history:
        return "No queries processed yet."

    latest = _history[-1]
    h = list(_history)
    n = len(h)

    avg_retrieval = safe_divide(sum(m["retrieval_ms"] for m in h), n)
    avg_generation = safe_divide(sum(m["generation_ms"] for m in h), n)
    avg_total = safe_divide(sum(m["total_ms"] for m in h), n)
    avg_tokens = safe_divide(sum(m["total_tokens"] for m in h), n)
    avg_eval = safe_divide(sum(m["eval_overall"] for m in h), n)
    fallback_pct = safe_divide(sum(1 for m in h if m["fallback"]), n) * 100

    return (
        f"**Latest Query**\n"
        f"- Retrieval latency: `{latest['retrieval_ms']:.0f}ms`\n"
        f"- Generation latency: `{latest['generation_ms']:.0f}ms`\n"
        f"- Total latency: `{latest['total_ms']:.0f}ms`\n"
        f"- Tokens used: `{latest['total_tokens']}`\n"
        f"- Eval score: `{latest['eval_overall']:.3f}`\n\n"
        f"**Rolling Average ({n} queries)**\n"
        f"- Avg retrieval: `{avg_retrieval:.0f}ms`\n"
        f"- Avg generation: `{avg_generation:.0f}ms`\n"
        f"- Avg total: `{avg_total:.0f}ms`\n"
        f"- Avg tokens/query: `{avg_tokens:.0f}`\n"
        f"- Avg eval score: `{avg_eval:.3f}`\n"
        f"- Fallback rate: `{fallback_pct:.1f}%`"
    )