app.py

import gradio as gr
import pandas as pd
import numpy as np
import matplotlib.pyplot as plt
from io import BytesIO
from datetime import datetime
import opik
import warnings
warnings.filterwarnings('ignore')

# --------------------------------------------------------
# CONFIG
# --------------------------------------------------------
OPIK_PROJECT_NAME = 'production-vf-ai'
OPIK_WORKSPACE_NAME = 'verba-tech-ninja'
OPIK_API_KEY = 'jYThN94NefoHKwEto3gPzhTSb'


# --------------------------------------------------------
# INIT OPik CLIENT
# --------------------------------------------------------
client = opik.Opik(
    api_key=OPIK_API_KEY,
    workspace=OPIK_WORKSPACE_NAME,
    project_name=OPIK_PROJECT_NAME
)


# --------------------------------------------------------
# FETCH TRACES
# --------------------------------------------------------
def fetch_traces(client_name, start_iso, end_iso):
    filter_string = (
        'name contains "analyse_transcript" '
        f'AND start_time >= "{start_iso}" '
        f'AND end_time <= "{end_iso}" '
        f'AND tags contains "{client_name}"'
    )

    traces = client.search_traces(
        project_name=OPIK_PROJECT_NAME,
        filter_string=filter_string,
        max_results=50000
    )
    return list(traces)


# --------------------------------------------------------
# FILTER TRACES
# --------------------------------------------------------
def filter_traces(traces):
    final = []
    for trace in traces:
        tags = trace.tags or []

        if "_call_" in tags or "[CAMPAIGN_CONVERSATION]" in tags:
            continue

        output = trace.output
        if not output:
            continue

        category = output.get("category")
        use_case = output.get("campaign_payload", {}).get("use_case")

        if category != "customer" and use_case is None:
            final.append(trace)

    return final


# --------------------------------------------------------
# PARSE SPANS
# --------------------------------------------------------
def extract_meta(trace):
    spans = client.search_spans(project_name=OPIK_PROJECT_NAME, trace_id=trace.id)
    out = []

    for s in spans:
        if s.name != "chat_completion_parse":
            continue

        usage = s.metadata.get("usage", {})
        out.append({
            "duration": s.duration / 1000,
            "tier": s.metadata.get("service_tier", "default"),
            "model": s.metadata.get("model"),
            "tokens": usage.get("completion_tokens", 0),
            "error": bool(s.error_info)
        })
    return out


# --------------------------------------------------------
# RUN MAIN PIPELINE
# --------------------------------------------------------
def run_pipeline(client_name, start_dt, end_dt, metadata_fields):

    start_iso = start_dt + "Z"
    end_iso = end_dt + "Z"

    traces = fetch_traces(client_name, start_iso, end_iso)
    traces = filter_traces(traces)

    rows = []
    for t in traces:
        rows.extend(extract_meta(t))

    if not rows:
        return "No data", None, None, None

    # Filter selected metadata fields
    df = pd.DataFrame(rows)
    df_filtered = df[metadata_fields]

    # ---------------- Stats -----------------
    durations = df.loc[~df["error"], "duration"]
    tokens = df["tokens"]

    stats = {
        "total_spans": len(df),
        "errors": int(df["error"].sum()),
        "error_rate_%": round(100 * df["error"].mean(), 2),

        "mean_latency_sec": round(durations.mean(), 3) if len(durations) else None,
        "median_latency_sec": round(durations.median(), 3) if len(durations) else None,
        "p90_latency_sec": round(durations.quantile(0.9), 3) if len(durations) else None,
        "p95_latency_sec": round(durations.quantile(0.95), 3) if len(durations) else None,
        "min_latency": round(durations.min(), 3) if len(durations) else None,
        "max_latency": round(durations.max(), 3) if len(durations) else None,

        "avg_tokens": round(tokens.mean(), 2),
        "max_tokens": int(tokens.max())
    }

    # ---------------- Charts -----------------
    fig1, ax1 = plt.subplots()
    ax1.hist(df["duration"], bins=30)
    ax1.set_title("Latency Distribution (seconds)")
    ax1.set_xlabel("Seconds")
    ax1.set_ylabel("Frequency")

    fig2, ax2 = plt.subplots()
    ax2.hist(df["tokens"], bins=25)
    ax2.set_title("Completion Token Distribution")
    ax2.set_xlabel("Tokens")
    ax2.set_ylabel("Frequency")

    # Convert figs to image
    buf1, buf2 = BytesIO(), BytesIO()
    fig1.savefig(buf1, format="png")
    fig2.savefig(buf2, format="png")
    buf1.seek(0)
    buf2.seek(0)
    plt.close(fig1)
    plt.close(fig2)

    # CSV
    csv_data = df_filtered.to_csv(index=False)

    return stats, df_filtered, buf1, buf2, csv_data


# --------------------------------------------------------
# GRADIO UI
# --------------------------------------------------------
with gr.Blocks(title="Opik Analytics Dashboard") as demo:

    gr.Markdown("# 📊 **Opik Analytics Dashboard** (Gradio)")
    gr.Markdown("Analyze traces by client, date range, and metadata fields.")

    with gr.Row():
        client_name = gr.Dropdown(
            ["fusiongroup", "vita", "staragent", "testclient", "other"],
            label="Select Client",
            value="fusiongroup"
        )

    with gr.Row():
        start_dt = gr.Textbox(label="Start DateTime UTC (YYYY-MM-DDTHH:MM:SS)", value="2025-11-17T00:00:00")
        end_dt = gr.Textbox(label="End DateTime UTC (YYYY-MM-DDTHH:MM:SS)", value="2025-11-17T12:00:00")

    metadata_fields = gr.CheckboxGroup(
        ["duration", "tier", "tokens", "model", "error"],
        label="Select Metadata Fields",
        value=["duration", "tier", "tokens"]
    )

    run_btn = gr.Button("Run Analysis")

    stats_output = gr.JSON(label="📈 Summary Statistics")
    table_output = gr.DataFrame(label="📄 Raw Data")
    plot_latency = gr.Image(label="⏱ Latency Distribution")
    plot_tokens = gr.Image(label="🔢 Token Distribution")

    csv_download = gr.File(label="⬇ Download CSV")

    run_btn.click(
        fn=run_pipeline,
        inputs=[client_name, start_dt, end_dt, metadata_fields],
        outputs=[stats_output, table_output, plot_latency, plot_tokens, csv_download]
    )


demo.launch()