Search with embedding vectors

Register the embedding model

Just like in the generation example, we initialize a Langdash instance and register an embedding model called “embedding_model”. Then we create an EmbeddingSearch object to search for embeddings using this model. We will use the all-MiniLM-L6-v2 served through the sentence_transformers backend.

ld = Langdash()
ld.register_model(
  name="embedding_model",
  model=Langdash.model_from_type(type="sentence_transformers", model="all-MiniLM-L6-v2"),
)

Search through the embeddings

Next, we create an embedding database, which stores a list of documents with their corresponding embeddings. The db.add() method is used to add multiple documents to the database.

db = EmbeddingSearch(ld.session_for_model("embedding_model"))

db.add(
  [
    "Oranges are delicious", "Apples are delicious",
    "Tokyo is the largest city in Japan", "The sky is blue",
    "An apple falls down from Tokyo Tower"
  ]
)

Finally, we use search(), which takes a query and a parameter max_documents which specifies how many candidate documents to return for each search (which defaults to 1). Here the example “where did the apple go” is searched. The index, the content and the similarity scores for each document are then printed.

query = "where did the apple go"
for i, doc, similarity in db.search(query):
  print(f"#{i} {doc}: {similarity}")
# example output: #2 An apple falls down from Tokyo Tower: 0.6126164197921753