Working with SPLADE v2

Learn how to use SPLADE with TVI.

What is splade?

SPLADE is similar to other inverted index approaches like BM25. SPLADE includes neural term expansion, meaning that it is able to match on synonyms much better than traditional BM25.

Using Splade with Trieve Vector Inference

1. Update embedding_models.yaml

   To use SPLADE with Trieve Vector Inference, you will need to adapt both the doc and query models

   The SPLADE document model is the model you use to encode files, where the query model is the one to encode the query that you will be searching with

   embedding_models.yaml

   models:
     # ...
     spladeDoc:
       replicas: 1
       modelName: naver/efficient-splade-VI-BT-large-doc
       isSplade: true
     spladeQuery:
       replicas: 1
       modelName: naver/efficient-splade-VI-BT-large-query
       isSplade: true
     # ...
   
2. Upgrade your TVI cluster

   Update TVI to include your models

   helm upgrade -i vector-inference \
       oci://registry-1.docker.io/trieve/embeddings-helm \
       -f embedding_models.yaml
   
3. Get embeddings endpoint

   kubectl get ing
   
4. Make call to generate sparse vector

   ENDPOINT="k8s-default-vectorin...elb.amazonaws.com"
   
   curl -X POST \
        -H "Content-Type: application/json"\
        -d '{"inputs": "test input"}' \
        --url http://$ENDPOINT/embed_sparse
   

   For more information checkout the API reference for sparse vectors.