mishelsikder’s mongo-predict repository trains a predictor from documents with known labels, then fills missing fields in other documents. An embedding model converts document text into numerical features; a fitted classifier or regressor produces the answer.

Labels change the comparison

The README reports AUROC 88.3 versus Jev’s 77.6 across 29 classification tasks, drawing on a separate Voyage study. AUROC measures how well scores rank classes across decision thresholds.

  • The supervised approach learns from each collection’s labeled examples; the Jev comparison uses zero-shot decisions.
  • The quoted result uses voyage-json-lite, while the package defaults to local voyage-4-nano embeddings and an XGBoost predictor.
  • Its “48% fewer errors” headline comes from comparing the complements of those AUROC scores. That is not a measured reduction in wrong labels.
  • The README’s token-price comparison leaves labeling, fitting and database operations outside the bill.

The project makes a concrete case for trying supervised learning when labels already exist. Its reported comparison cannot establish the same advantage on a new, unlabeled workload.

The prediction stage runs in Python

$predict is implemented by a PyMongo adapter, with ordinary aggregation continuing on MongoDB:

  1. Fetch the selected documents and compute predictions on the client.
  2. Upload prediction rows to a temporary collection.
  3. Join them to source documents and run the remaining pipeline.
  4. Remove the temporary collection when the call finishes.

That means client compute, database write permissions and data movement belong in the deployment plan. Stored models use Python pickle; restrict who can write model records and consider the supported signing option.

The movie demo has a clearer result

The authors’ saved experiment trains on 17,527 pre-2010 movies and tests 5,914 newer movies. It reports mean absolute rating error of 0.675, versus 0.764 for always guessing the older films’ average.

Its recorded fitting and prediction times exclude fresh embedding work. Neither the movie result nor the Jev comparison was independently reproduced here.

Caveat: the candidate repository’s install command points to a different GitHub owner, and its package metadata credits Chandan Singh. Confirm package provenance before installing.

Verification: inspected the complete README, prediction and model-management code, embedding definitions, movie demo and saved results. Recomputed the headline arithmetic only; no database workload, model inference or Jev evaluation was run.