"""Vega-Lite specs for the daily-granularity charts.""" import json import statistics from datetime import date from chart_style import BG, CONFIG, EMPTY, PLOT_WIDTHS, RAMP, SPEND, TOKENS, publish def spec(**kw): s = {"$schema": "https://vega.github.io/schema/vega-lite/v6.json", "config": CONFIG} s.update(kw) return s cal = json.load(open("calendar.json")) cum = json.load(open("cumulative.json")) per_model = json.load(open("per_model_daily.json")) model_order = json.load(open("model_order.json")) when = json.load(open("when.json")) def pretty(day: str) -> str: d = date.fromisoformat(day) return f"{d.day} {d.strftime('%B %Y')}" span_text = f"{pretty(cal[0]['date'])} to {pretty(cal[-1]['date'])}" n_weekday = sum(1 for r in cal if r["weekday"] < 5) n_weekend = len(cal) - n_weekday n_weeks = len({r["week"] for r in cal}) # --- 1. calendar heatmap ----------------------------------------------------- active = sorted(r["millions"] for r in cal if r["tokens"]) cuts = statistics.quantiles(active, n=4) print("heatmap thresholds (millions):", cuts) def mag(value: float) -> str: """Enough precision to name the bin. The quietest quartile is a fraction of a million a day --- the Zed years --- and 'under 0M' means nothing.""" return f"{value:.1f}".rstrip("0").rstrip(".") if value < 10 else f"{value:.0f}" BINS = [ "none", f"under {mag(cuts[0])}M", f"{mag(cuts[0])}-{mag(cuts[1])}M", f"{mag(cuts[1])}-{mag(cuts[2])}M", f"over {mag(cuts[2])}M", ] calendar = spec( title={ "text": "Every day of agent use", "subtitle": f"Tokens per day, {span_text}. " "Unfilled cells are days with no session at all.", }, data={"values": cal}, transform=[ { "calculate": f"datum.tokens === 0 ? '{BINS[0]}'" f" : datum.millions < {cuts[0]} ? '{BINS[1]}'" f" : datum.millions < {cuts[1]} ? '{BINS[2]}'" f" : datum.millions < {cuts[2]} ? '{BINS[3]}' : '{BINS[4]}'", "as": "bin", } ], width=PLOT_WIDTHS["daily-heatmap"], # square cells, whatever the span: seven rows of the column width height=round(7 * PLOT_WIDTHS["daily-heatmap"] / n_weeks), mark={"type": "rect", "cornerRadius": 2, "stroke": BG, "strokeWidth": 2}, encoding={ "x": { "field": "date", "type": "ordinal", "timeUnit": "yearweek", "axis": { "title": None, "labelAngle": 0, "grid": False, "domain": False, "ticks": False, "labelOverlap": False, "labelExpr": "date(datum.value) > 7 ? '' : month(datum.value) == 0" " ? timeFormat(datum.value, '%b %y') : timeFormat(datum.value, '%b')", }, }, "y": { "field": "date", "type": "ordinal", "timeUnit": "day", "axis": { "title": None, "grid": False, "domain": False, "ticks": False, "format": "%a", "labelExpr": "indexof([1,3,5], day(datum.value)) >= 0 ? timeFormat(datum.value, '%a') : ''", }, }, "color": { "field": "bin", "type": "nominal", "scale": {"domain": BINS, "range": [EMPTY] + RAMP}, "legend": { "title": None, "orient": "bottom", "direction": "horizontal", "symbolType": "square", "symbolSize": 130, "columnPadding": 12, }, }, "tooltip": [ {"field": "date", "type": "temporal", "title": "day"}, { "field": "millions", "type": "quantitative", "title": "million tokens", "format": ",.0f", }, ], }, ) # --- 2. small multiples: daily tokens per model ------------------------------ by_model_daily = spec( title={ "text": "The rise and fall of each model", "subtitle": "Tokens per day, 7-day rolling mean. One panel per model, ordered by first use; " "all panels share the same scale.", }, data={"values": per_model}, transform=[ { "window": [{"op": "mean", "field": "millions", "as": "smooth"}], "frame": [-3, 3], "groupby": ["model"], "sort": [{"field": "date"}], }, ], facet={ "row": { "field": "model", "type": "nominal", "sort": model_order, "title": None, "header": { "labelAngle": 0, "labelAlign": "left", "labelOrient": "left", "labelPadding": 6, }, } }, spec={ "width": PLOT_WIDTHS["rise-and-fall"], "height": 36, "mark": { "type": "area", "color": TOKENS, "opacity": 0.9, "interpolate": "monotone", "line": False, }, "encoding": { "x": { "field": "date", "type": "temporal", "axis": { "title": None, "format": "%b %y", "labelAngle": 0, "grid": False, "tickCount": {"interval": "month", "step": 2}, }, }, "y": { "field": "smooth", "type": "quantitative", "axis": { "title": None, "grid": False, "labels": False, "ticks": False, "domain": False, }, }, "tooltip": [ {"field": "model", "type": "nominal"}, {"field": "date", "type": "temporal", "title": "day"}, { "field": "smooth", "type": "quantitative", "format": ",.0f", "title": "million tokens/day", }, ], }, }, resolve={"scale": {"y": "shared"}}, spacing=3, ) # --- 3. cumulative ----------------------------------------------------------- cum_rows = [] for r in cum: cum_rows.append( {"date": r["date"], "measure": "billions of tokens", "value": r["billions"]} ) cum_rows.append( { "date": r["date"], "measure": "indicative API cost (US$)", "value": r["dollars"], } ) cumulative = spec( title={ "text": "Cumulative since day zero", "subtitle": f"The same {len(cal)} days, counted two ways." " Separate panels, separate scales.", }, data={"values": cum_rows}, facet={ "row": { "field": "measure", "type": "nominal", "sort": ["billions of tokens", "indicative API cost (US$)"], "title": None, "header": { "labelAngle": 0, "labelAlign": "left", "labelOrient": "top", "labelPadding": 2, }, } }, resolve={"scale": {"y": "independent"}}, spec={ "width": PLOT_WIDTHS["cumulative"], "height": 160, "mark": {"type": "area", "opacity": 0.85, "line": {"strokeWidth": 2}}, "encoding": { "x": { "field": "date", "type": "temporal", "axis": { "title": None, "format": "%b %y", "labelAngle": 0, "grid": False, "tickCount": {"interval": "month", "step": 2}, }, }, "y": { "field": "value", "type": "quantitative", "axis": {"title": None, "format": "~s"}, }, "color": { "field": "measure", "type": "nominal", "legend": None, "scale": { "domain": ["billions of tokens", "indicative API cost (US$)"], "range": [TOKENS, SPEND], }, }, "tooltip": [ {"field": "date", "type": "temporal", "title": "day"}, {"field": "value", "type": "quantitative", "format": ",.0f"}, ], }, }, ) # --- 4. diurnal profile ------------------------------------------------------- diurnal = json.load(open("diurnal.json")) when_chart = spec( title={ "text": "The working day, as seen by the token meter", "subtitle": "Average tokens per hour of the day (Australia/Sydney), across" f" {n_weekday} weekdays and {n_weekend} weekend days.", }, data={"values": diurnal}, width=PLOT_WIDTHS["diurnal"], height=250, layer=[ { "mark": { "type": "line", "strokeWidth": 2, "interpolate": "monotone", "point": {"filled": True, "size": 45}, }, "encoding": { "tooltip": [ {"field": "kind", "type": "nominal", "title": None}, {"field": "hour", "type": "quantitative"}, { "field": "millions", "type": "quantitative", "format": ",.1f", "title": "million tokens/hour", }, ] }, }, { "transform": [ { "filter": "(datum.kind === 'Mon-Fri' && datum.hour === 17) || (datum.kind === 'Sat-Sun' && datum.hour === 21)" } ], "mark": { "type": "text", "align": "left", "dx": 9, "fontSize": 13, "fontWeight": 600, }, "encoding": {"text": {"field": "kind", "type": "nominal"}}, }, ], encoding={ "x": { "field": "hour", "type": "quantitative", "axis": { "title": None, "grid": False, "values": [0, 3, 6, 9, 12, 15, 18, 21], "labelExpr": "datum.value + ':00'", }, "scale": {"domain": [0, 23], "nice": False}, }, "y": { "field": "millions", "type": "quantitative", "axis": {"title": "million tokens per hour"}, }, "color": { "field": "kind", "type": "nominal", "legend": None, "scale": {"domain": ["Mon-Fri", "Sat-Sun"], "range": [TOKENS, SPEND]}, }, }, ) for name, s in ( ("daily-heatmap", calendar), ("rise-and-fall", by_model_daily), ("cumulative", cumulative), ("diurnal", when_chart), ): publish(name, s)