Atlas creates Charts calculations and plots them

Summary

Let Atlas create a Chart calculation from a natural-language request and plot the result as a time series chart.

Today a user can build that calculation by hand in the Charts node editor (sources, constants, functions, and an output node), or ask Atlas to analyze stored time series. Atlas can embed a bar chart, KPI, table, or an existing Charts time series in a reply. It cannot turn a formula, for example a pressure difference, a unit conversion, or a scaled flow, into a calculation and show that calculated series on a line chart unless the result already exists as a time series in CDF.

Problem

Engineers and operators compare derived signals as often as raw tags: pressure drop, API gravity, unit conversion, ratio of two flows, a moving average. Those series are usually not stored in CDF.

The path today is manual:

1. Find the source time series.

2. Open Flows > Charts, create a chart, and add the series.

3. Open Add data > Create calculation and wire sources, constants, and functions into an output node.

4. Read the plot, then adjust the calculation if the result is wrong.

Atlas already sits next to this work. In Charts, Agents can find data and discuss thresholds. In Ask Atlas AI, the agent can retrieve time series, run Analyze time series, or compute a conversion in the sandbox. The calculated values stay in the answer. The line chart still shows only stored series, so the user cannot see the derived trend against the raw tags.

That gap shows up when:

• Two series are comparable only after a conversion (different units or a scale factor).

• The user wants several one-off calculations (difference, ratio, smoothed signal) without a library of persisted calculated time series.

• The user knows the engineering formula and the asset, and does not want to assemble the node graph.

Current behavior

Proposed solution

Add an Atlas capability that:

1. Understands a request to derive and plot a time series.

2. Resolves the source time series in CDF (including @ mentions and the active location filter).

3. Builds a Charts calculation using the same primitives the node editor already supports: sources, constants, functions, and an output.

4. Evaluates the calculation with the Charts calculation engine, not with the language model.

5. Renders the output as a time series line chart, with the source series available on the same chart when that helps comparison.

6. Lets the user open the result in Charts, inspect and edit the node graph, save the chart and only after an explicit confirmation; schedule it as a time series.

The calculation definition must be inspectable. The user sees which series, constants, and functions Atlas used before treating the plot as correct.

Example requests

• “Plot the pressure difference between PT-1019 and PT-1017 for the last 7 days.”

• “Convert this flow from tons per period to tons per hour and chart it next to the raw series.”

• “Show API gravity from @specific-gravity using 141.5 / SG − 131.5.”

• “Add a 1-hour moving average of bearing temperature on compressor C-205 to this chart.”

User story

As a reliability or process engineer using Cognite Data Fusion,
I want Atlas to create a Charts calculation from my request and plot the result,
so that I can see a derived trend (difference, conversion, ratio, or smoothed signal) without building the calculation by hand or storing a new time series first.

Requirements

Discovery and intent

1. Atlas recognizes requests whose goal is a plotted derivation, including difference, sum, ratio, scale, unit conversion by a constant, and a function that Charts already exposes (for example a moving average).

2. A request that only needs a number or a statistical conclusion continues to use the existing analysis tools. Atlas plots a calculation when the user asks to chart, plot, trend, compare visually, or add a series to a chart.

3. The user can make the request from:

   • Ask Atlas AI in the Flows workspace

   • the Agents panel on an open chart

4. @ mentions of a time series, asset, or activity are treated as pinned inputs and take precedence over name search.

5. Location filters selected in the agent session limit which time series Atlas may use.

Building the calculation

• Atlas resolves each input to a specific CDF time series (external ID or instance identity) before it builds the calculation.
• If more than one time series matches, or a required input, constant, unit, or time range is missing, Atlas asks a clarifying question and does not plot a guess.
• The calculation is a valid Charts calculation: one or more source nodes, optional constant nodes, function nodes supported by Charts, and exactly one output node for each plotted result.
• Arithmetic and function evaluation run in the Charts calculation engine (or the same engine Charts uses). The language model must not invent data points.
• The calculation records the output name, the formula in user language, the node graph, each source time series, each constant, and the time range used to render.
• When the request depends on aligning series with different timestamps, Atlas uses Charts automatic data alignment and states that it did so.
• A calculation that Charts cannot express (custom packages, a function that is not in the Charts library, or a formula that needs Cognite Functions) is declined with a short reason. Atlas does not approximate it with model-generated points.

Chart output

• The output is a time series line chart of the calculation result over the requested range, or over the chart’s current range when the user is in Charts and does not name a range.

• Source series used in the calculation are available on the same chart so the user can compare raw and derived traces. The user can hide a source series without deleting the calculation.

• The chart shows the calculation name, unit when known, and a visible indication that the trace is calculated rather than a stored time series.

• In Ask Atlas AI, the chart is rendered in the reply and can be opened in Charts.

• In the Charts agent, the calculation is added to the open chart and appears both on the plot and in the calculation editor.

• The user can change the time range and the calculation re-evaluates for that range, the same way a manual Charts calculation does.

• Empty inputs, a divide-by-zero, or a range with no points produce an explicit message and an empty or partial plot. The chart does not show a fabricated line.

Review, save, and schedule

• Before or alongside the plot, Atlas shows the series, constants, and functions it used, in language the user can check (for example “PT-1019 − PT-1017”).

• The user can open the calculation in the Charts node editor and change it. Edits update the plot.

• Saving the chart follows the existing Charts save behavior and permissions.

• Scheduling (Save and Schedule) and writing a new CDF time series happen only after the user explicitly confirms. Atlas pre-fills the calculation; it does not start a schedule on its own.

• One-off calculations that the user does not save do not create time series, schedules, or monitoring jobs.

Access and safety

• Atlas can read only time series the signed-in user can read, and can create or update a chart only when that user can do so in Charts.

• The feature respects the same Atlas AI capabilities required to chat with an agent.

• Citations or source references identify the input time series so the user can open them in Search or Charts and verify the inputs.

Quality of the result

• For the in-scope formulas below, the plotted values match the same calculation built manually in Charts on the same sources, constants, alignment setting, and time range.

• In-scope formulas for the first release:

  • add, subtract, multiply, and divide two series or a series and a constant

    • a constant scale or offset, including a stated unit conversion that is a single factor

    • a chained expression that Charts can represent as ordered nodes (for example 141.5 / SG − 131.5)

    • one Charts smoothing or resampling function if that function already exists in the calculation editor (confirm the exact function list with the Charts team)

• Atlas states the time range, alignment, and unit assumptions in the reply.

In scope

• Natural-language creation of a Charts calculation and its line chart

• Ask Atlas AI and the agent panel inside Charts

• Inputs by name, by asset context, and by @ mention

• Clarifying questions when the series, constant, or range is ambiguous

• Inspectable node graph, open-in-Charts, and manual edit

• Explicit, user-confirmed save and schedule using the current Charts flows

• Parity with a manually built Charts calculation for the formulas 

Out of scope

• Persisting a calculated time series or starting a schedule without confirmation

• New calculation functions that Charts does not already support

• Cognite Functions, custom Python packages, or model-written code as the plotting engine

• Bulk generation of large chart libraries

• Monitoring and alerting on the new calculation, beyond the existing Charts monitoring flow the user starts themselves

• Bar charts, KPI tiles, and tables of non-time-series aggregates (already available)

• Animating or streaming the series beyond the live behavior Charts already has

• Changing who can read a time series