Skip to main content

Configuring models to ingest features from other models

Sometimes the outputs of one model should be "chained together" to become inputs of another model. Typical examples of this are:

  • Chunking and vectorizing text data, to become input to a retrieval augmented LLM.
  • Transfer learning in computer vision, where prepared features become input to a shallow classifier

Procedural API

In procedural mode, pass the output of one .predict as a dependency of the second, as well as specifying that the model should ingest the outputs of the first, using the .outputs query.

j1 = m1.predict(X='my_key_1', select=collection.find())
select=collection.find().outputs('my_key_1', m1.identifier),

In CFG.production = True mode, the j1 object is a Job object, which doesn't block the main thread of the program. The second .predict waits for this Job to finish before starting.

Declarative API

With the declarative API this behaviour may be achieved by linking the .outputs of one Listener as the select=... of another:

l1 = Listener(

l2 = Listener(

This implies that whenever data is inserted to collection, model_1 will compute outputs on that data first, which will subsequently be consumed by model_2 as inputs; its outputs will then also be saved to db.