Elai Platform v 2.41.2
Your AI agent can read now
Minor release, November 2026
Welcome to the new release! This one is a big one for the AI agent chat. It finally learned to read your documents, it now hangs out with you in the tagging step (where it can even write SQL for you), and it knows a thing or two about clusters. But the agent is not the only one getting attention: hierarchical forecasting (HTS) gets a brand new Evaluate model dashboard and a smart helper that builds your hierarchy for you, Forecast models with an Item ID can finally run predictions, and editing a model got a lot quicker.
Grab a coffee, here is everything that is new.
- Talk to your documents
- The AI agent joins the tagging step
- Ask the agent about your clusters
- A brand new Evaluate model dashboard for HTS
- Let us build the hierarchy for you
- Predictions for Forecast models with an Item ID
- Small edits, small effort
- Leave unwanted records at the source
- Other noteworthy changes
- Under the hood
📄 Talk to your documents
Until now, the AI agent only knew what it could find in your data and what you typed in the chat. That is a lot, but often the really useful context lives somewhere else: a business requirements document, a spreadsheet with the product catalog, a slide deck from the last steering meeting.
Starting with this release, you can attach files directly to a chat message. Pick one or more files, write your question, and the agent will take them into account when answering. Supported formats are MD, DOCX, PDF, CSV, TXT, XLSX, JSON, Parquet and PPTX, up to 20 MB per file.
Documents can also live at the model level. The model context page has been reorganized into two sections:
- Memory (formerly “Context”): the free text notes the agent keeps about the model, as before.
- Knowledge: the files you uploaded for this model. These are always available to the agent in every conversation about this model, so you do not have to attach the same document over and over.
Processing a document takes a little time. While the agent is reading it, the page shows its status and updates by itself as soon as the file is ready, no refresh needed.
🏷️ The AI agent joins the tagging step
Tagging a dataset is one of the moments where a second pair of eyes helps the most, so the agent now follows you there too.
The chat is aware of the table you are currently looking at: when you switch table or group, the agent’s context follows along. To get you started, there is a new Dataset overview suggestion that asks the agent to describe the current dataset: its main characteristics, the kind of data it contains and the most relevant figures. From there, you can keep asking about specific columns, values or how they have been tagged.
And when a question is better answered with a query, the agent can now write SQL queries on the table you are working on. Queries show up as proper code blocks, now with syntax highlighting, so they are easy to read and ready to copy.
🔍 Ask the agent about your clusters
Clustering models are great at finding groups, but a scatter plot does not always tell you why those groups exist. The AI agent chat has a new Identified clusters suggestion: it explains the clusters found by the model and what makes each of them different from the others.
📊 A brand new Evaluate model dashboard for HTS
Hierarchical forecasting models get a completely redesigned Evaluate model page, built around the hierarchy itself.
You can now:
- Browse the hierarchy level by level, from the most aggregated series down to the bottom-level ones.
- Search for a specific node within a level, so you do not have to scroll through thousands of stores or products.
- Select multiple nodes at once and compare them.
- See the distribution of each node, that is the share of dataset records that belong to it, and sort nodes by it.
- Open the results of every single time series, including the actual values over the prediction period for bottom-level series too.
A heads up: the new dashboard relies on a new way of storing results. HTS models trained with a previous version will ask you to retrain them before showing the new dashboard.
🌳 Let us build the hierarchy for you
Setting up an HTS model means putting your ID columns in the right order. If you tagged product, store and category, they need to become store, then category, then product. Easy with three columns, less so with ten.
Now, when you enable hierarchical aggregation in the model settings, the platform analyzes your data and suggests a hierarchy, filling it in for you. It recognizes columns that nest into each other, columns that are constant, and columns that cross the hierarchy (think gender, or sizes that depend on the department) and places each of them in a sensible spot.
The analysis can take a few seconds on large datasets: a loading state keeps you posted and you can cancel it at any time to configure the hierarchy manually. Changed your mind after editing? Hit Recalculate hierarchy to start again from the suggestion. And of course, if you already defined a hierarchy, it is left untouched.
📈 Predictions for Forecast models with an Item ID
Forecast models trained with an Item ID column could already learn from many products at once, but running predictions on them was not possible yet. This release closes the gap: the Item ID tag is now available in prediction datasets for Forecast models that used it in training, and the whole prediction pipeline knows how to map each product back to the right series.
Speaking of Forecast models, the health check now includes the Target summary for them too, both at group and global level, whether the model is single target, multi target, with an Item ID or HTS.
✅ Small edits, small effort
We have all been there: you just want to ignore one column in a multi-table model with time windows, and suddenly you are clicking through every single step to confirm. Not anymore. Every step of the tagging flow now has its own Confirm button, so you can make your change and save right away. The usual rules still apply: the button is enabled only when you actually changed something and all steps are valid.
Also new in the data loading area: Append and Replace now work with the Repository as a source. Pick an existing file from the repository and add it to, or swap it into, an existing dataset, exactly as you would with a fresh upload.
🚫 Leave unwanted records at the source
When importing from a SQL connector, you can now add the Exclude records… operation already in the preview phase. The filter is applied directly by the connector, based on the relations between your tables, so records you do not need are never transferred in the first place. Less data to move, faster imports.
In the operations panel, these operations carry an extra icon so you can tell at a glance which ones run on the connector side.
🔮 Say hello to Timestral
Forecast models have a new algorithm in the race: Timestral. It is a family of probabilistic, multivariate forecasting models that can also take into account contextual variables, both past and future.
What makes it different from the classic algorithms is how it learns. Timestral models are pre-trained on large amounts of data and then fine-tuned on real-world data. This lets them pick up general patterns and deliver accurate forecasts even on datasets they have never seen before, without a dedicated training (the so-called zero-shot forecasting).
Timestral competes with the other algorithms during training like any other candidate: when it wins, you will find it in the model results with its own name and description.
✨ Other noteworthy changes
- In the Analyze scenarios dashboard you can now download the prediction file for just the product you are looking at, instead of always getting all of them. The download menu offers both options: the selected product, or all products at once.
- The feature selection agent gained a new step: numeric and categorical column combination. It creates new features by splitting a numeric column across the most frequent categories of a categorical one (for example, total payment amount per card type). You can turn it on or off and cap the maximum number of combinations from the agent configuration.
- The dataset card now shows a progress percentage for the analysis phase, alongside prediction and test, so you know how far along it is instead of staring at a spinner.
- We said goodbye to the old predefined, multi-language table names. For a few releases now, new tables take the name of their dataset and can be renamed at any time. If some of your tables still used the old names, they keep working and will now simply show their English version (Subjects, Payments, and so on). As always, you can rename them whenever you like.
🔧 Under the hood
Not everything in a release is something you can click on. A few changes that make the platform sturdier:
- Smoother feature selection under load. Feature selection can now spread its work across more resources, so it keeps up when many models are being trained at the same time. A big model no longer slows everyone else down.
- Smarter imports from connectors. When many large imports arrive at once, the platform now lines them up and processes them a few at a time instead of all together. This keeps everything stable, even on busy days.
- Health checks for Forecast models. Health checks gained new controls dedicated to Forecast models, including models with an Item ID and global targets.
- Faster unique values. When nothing has changed in a dataset, unique values are read from the already processed file instead of the original one, which is noticeably faster on large datasets.
- Row-based target builder for Forecast models. For Forecast models, the target builder now always works row by row and no longer groups records by primary IDs.