> For the complete documentation index, see [llms.txt](https://docs.beye.ai/llms.txt). Markdown versions of documentation pages are available by appending `.md` to page URLs; this page is available as [Markdown](https://docs.beye.ai/product-docs/analytics-harness.md).

# Analytics Harness

## Analytics Harness

One proprietary engine behind every answer, dashboard and calculator in B/eye.

B/eye is an analytics harness designed for supply chain use cases in the enterprise, across warehouse distribution, 3PL and manufacturing. There is a single harness, and it runs everywhere. The same harness does the work whether you are asking an ad hoc question, building a report or dashboard to store in a stack, or running a calculator for advanced analysis. There are no modes to choose between. You ask in plain language and the harness handles the rest.

### What the Harness Does for Every Question

For each question, the harness:

1. Works out where the data lives, in one system or several.
2. Cleanses the data and brings it together.
3. Builds the queries and analysis with SQL and Python.
4. Validates the results, with checks on joins, filters, time windows and totals.
5. Creates the reporting: figures, tables, charts and written summaries.

That gives operators, inventory teams and other business colleagues their time back for decisions, instead of spending it preparing to do the work.

### Ask a Question

1. Open **Conversational Analytics** from the left rail.
2. In the sidebar, pick the dataverse you want to work with, then open a conversation or start a new one.
3. Type your question in **Ask a question...** and send it.
4. Follow up in plain language to refine the answer. Follow ups remember context, so you can narrow by region, DC, shift or account without starting over.

<figure><img src="https://3313864313-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzL3krDaq8tE17VN9J3mX%2Fuploads%2FKF7aSxTCFUZ8O8jWRgzW%2Fad-hoc-analysis.png?alt=media&amp;token=bdc4bf5a-84e0-49a1-87ac-16e8d3dfa903" alt=""><figcaption></figcaption></figure>

*An ad hoc question answered with Python. Every step the harness took is visible.*

### See Exactly How Each Answer Was Built

Nothing is hidden behind a spinner. As the harness works, each step appears in the conversation:

* **Thought** explains what the harness is about to do and why.
* **Looked up metadata** shows the harness reading your dataverse metadata to find the right tables and columns.
* **Analyzed data with code** shows the harness writing and running SQL and Python, with how many queries ran and how long it took.

Open an analysis step to review the work:

| Tab         | What you see                                  |
| ----------- | --------------------------------------------- |
| **PYTHON**  | The exact code that produced the answer       |
| **QUERIES** | The SQL queries the code ran                  |
| **RESULT**  | The output                                    |
| **STATUS**  | Progress notes written while the analysis ran |

Use the copy button to copy the code. Each result shows its title, the dataverse it came from, the row count and the run time.

### What You Can Get Back

* **Summary figures** for the headline numbers.
* **Full tables** you can sort, filter by column, browse and download.
* **Charts**, including scatter plots on numeric axes and points colored by a category such as customer segment, region or product line. When the harness clusters your data, it marks each cluster's center so the segments are visible at a glance.
* **Several results at once.** A single analysis can return up to five results, mixing figures, tables and charts. A question such as "segment these customers and show me how the groups differ" can come back as the segmentation chart, the per segment summary and the underlying table together.
* **PDF reports.** An analysis can produce a branded, downloadable PDF that combines written commentary, summary cards, status checks, tables and charts. Charts in the PDF look exactly like the ones on screen.
* **Categorized free text.** Give the harness a fixed set of labels, such as billing, shipping and product quality, and it assigns one to every row in a text column. Because the labels are set up front, results stay consistent across the whole column. Up to 2,000 rows can be categorized in a single run.

### Advanced Analysis With Python

When SQL alone cannot reach the answer, the harness writes and runs its own Python. It uses the established data science toolkit, including NumPy, SciPy, scikit-learn, XGBoost and LightGBM, so answers come from proven methods rather than approximations.

* Forecast a trend forward
* Detect anomalies and outliers
* Cluster customers or products into segments
* Fit regressions and test correlations
* Train and compare predictive models

Each analysis runs in an isolated environment with no network access and can only reach the data in your dataverse.

### Choose a Model and Reasoning Depth

The harness works the same way no matter which model you pick. The model picker sits under the question box and shows the current model and reasoning depth.

* **Search models...** to find a model by name, or filter the list by provider.
* Select the star next to a model to add it to your favorites.
* Adjust the reasoning depth for each model, from low through the maximum setting.
* Any default you have set stays in place, and existing conversations keep the model they were started with.

<figure><img src="https://3313864313-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzL3krDaq8tE17VN9J3mX%2Fuploads%2F9UtePtr2zeLvj7sICgbq%2Fmodel-picker.png?alt=media&amp;token=cf2bb41d-680c-4c27-985b-d0afd9a789ed" alt=""><figcaption></figcaption></figure>

*The model picker. Choose a model and set how deeply it reasons.*

### Dig Into Any Figure

Hover over any figure on a dashboard, whether it is a KPI value, a chart or a table, to see what produced it: the query behind it, the value and its format, the filters in effect and any caveats that apply.

Select **Dig into this figure** to open a new question with that figure's query, value and active filters already attached. The harness checks the number against the underlying data and explains how it was derived.

<figure><img src="https://3313864313-files.gitbook.io/~/files/v0/b/gitbook-x-prod.appspot.com/o/spaces%2FzL3krDaq8tE17VN9J3mX%2Fuploads%2FanP6Jk4vifWubzE2RJ72%2Fdig-into-figure.png?alt=media&amp;token=05c69391-4f33-4e86-b571-a976a13939b4" alt=""><figcaption></figcaption></figure>

*Every figure shows the query behind it. Dig in to ask how the number was derived.*

Digging in is available to anyone with access to the underlying data. Readers of a published report can still see the query details behind each figure.

### Questions to Start With

* "Where are we most at risk of stockouts this week by SKU and retailer, and what coverage do we have at the nearest DCs?"
* "What is today's stockout rate by SKU and location, what is the revenue impact, and which nodes account for most of the gap?"
* "Which suppliers and lanes are missing on time in full, and how does that affect fill rate and cost to serve?"
* "Where do we have low days of supply, and which safety stock or lead time change removes the most risk with the least added working capital?"
* "Which business units drive 80% of the variance?"

### Tips for Better Answers

* **Be specific.** Name the metric, the time range and how you want it grouped.
* **Ground your definitions.** For example: "Use landed cost = base + tariff + freight."
* **Start broad, then narrow down.** Each follow up refines the result in place.
* **Ask for assumptions.** "List the business rules you applied and show the SQL you generated."
* **Keep one topic per conversation** so the context stays clean.
* **Define measures once.** Add recurring calculations as measures in your dataverse instead of re-deriving them in chat. B/eye reuses them automatically.
* **Review before you share.** AI can make mistakes, so verify important numbers. Every query and line of code is there for you to check.