Voyage AI: Embeddings Cut Prediction Errors on MongoDB Data by Up to 44%

October 28, 2026, 01:00Source: Voyage AI

Voyage AI says its embedding models can predict missing fields in MongoDB collections more accurately than TypeSafe AI's Jev. Across 50 real-world prediction tasks, the company reports that adding Voyage embeddings of the whole document reduces errors by 37% with voyage-4-nano and by 44% with voyage-json.

Key points

Voyage AI says adding its embeddings to raw fields cuts errors by 37% (voyage-4-nano) and 44% (voyage-json) across 50 prediction tasks.
On 29 classification tasks, voyage-json scores 88.3 AUROC versus 77.6 for TypeSafe AI's Jev, or 48% fewer errors.
On movies from 2010 on, the model had an average error of 0.68 rating points versus 0.76 for guessing the average.
The method embeds JSON documents and fits a model such as XGBoost; the community library mongo-predict wraps the steps.
voyage-json is experimental; voyage-4-nano is open-weight and runs locally.

Learning from documents already in the collection

Many MongoDB collections contain a field that is set in some documents and missing in others. Voyage AI gives the example of a support ticket's priority, which may only be set on older tickets. MongoDB documents also mix scalar fields with free text, lists and nested sub-documents. According to the company, tabular models need a flat table, so the usual approach is to flatten each document and drop what does not fit, which hurts accuracy.

General models like Jev label documents directly, without learning from the labeled documents already in the collection. Voyage AI describes a way to build a version that does learn from them. The steps are:

Turn each document into a JSON string, leaving out the field to predict.
Embed each JSON string with a Voyage model, for example voyage-4-nano locally or voyage-4-large through the Voyage API.
Fit a prediction model such as XGBoost on the embeddings of documents where the field is set.
Predict the field for the remaining documents and write it back to the collection.

Because the embeddings only need to be computed once, testing different prediction models is fast, according to the post.

Results on movies and 50 tasks

For a worked example, Voyage AI used MongoDB's sample_mflix.movies collection. It fitted a model on the 17,527 movies released before 2010 and predicted the IMDb rating of the 5,914 movies released from 2010 on. The model used title, year, runtime, age rating, genres, cast, directors and plot. Other ratings in the documents, such as Rotten Tomatoes and Metacritic, were not used.

The model had an average error of 0.68 rating points and predicted 78% of movies within one point of the actual rating. Always guessing the average rating of the older movies gave 0.76 and 72%. Once embeddings were computed, fitting took 79 seconds and predicting all 5,914 ratings took 8 seconds on 16 CPU threads.

The broader evaluation covers 50 tasks: 7 from MongoDB's sample data, all 21 tasks of TextTabBench and 22 collected from the web. Of these, 29 predict a category and 21 predict a number. The raw fields alone score 52.1 on average. Adding voyage-json lifts the average to 73.3 and voyage-4-nano to 70.0, which Voyage AI says is 44% and 37% fewer errors. The six open baseline models score between 62.9 and 70.8, or 23% to 39% fewer errors. According to the company, voyage-4-nano outperforms all of them except Nemotron-3-Embed-8B, which has over 20 times as many parameters.

Comparison with Jev and available tooling

On the 29 classification tasks, voyage-json reaches an average AUROC of 88.3, compared with 77.6 for TypeSafe AI's Jev, which Voyage AI says is 48% fewer errors. The post notes that, unlike Jev, this approach needs labeled documents. voyage-json is an experimental model that the company post-trained from voyage-4-lite for predictions on JSON documents.

voyage-4-nano is an open-weight model that runs locally, on a GPU if available. voyage-4-lite, voyage-4 and voyage-4-large are available through the Voyage API and need no GPU. Any scikit-learn estimator can be fit on the embeddings. The evaluation used XGBoost for movie ratings and ridge regression on the raw fields plus the embedding for the 50 tasks.

Voyage AI also points to mongo-predict, a community library that wraps these steps for MongoDB. It adds helper methods to pymongo collections that fit a prediction model and store it in the database. A $predict step at the start of an aggregation pipeline embeds and scores the matching documents in the Python client, while the rest of the pipeline runs on the server. The library is open source and available on GitHub, and uses voyage-4-nano as its default embedding model.

What it means for MongoDB users

For teams with partly labeled collections, the post presents a route to fill gaps without flattening documents into tables. The figures are Voyage AI's own evaluation, and the approach requires existing labeled documents in the collection. The company says the same method covers imputation, classification and regression.