Exemple de modèle d’incorporation fondamental de recherche vectorielle (GTE)

Ce notebook montre comment utiliser le SDK Python de recherche vectorielle, qui fournit un VectorSearchClient en tant qu'API principale pour travailler avec la recherche vectorielle.

Ce notebook utilise les API Databricks Foundation Model pour accéder au modèle d'embeddings GTE afin de générer des embeddings.

%pip install --upgrade --force-reinstall databricks-vectorsearch
dbutils.library.restartPython()

from databricks.vector_search.client import VectorSearchClient

vsc = VectorSearchClient(disable_notice=True)

help(VectorSearchClient)

Charger le jeu de données d'exemple dans une table Delta source

L’exemple suivant crée la table Delta source.

# Specify the catalog and schema to use. You must have USE_CATALOG privilege on the catalog and USE_SCHEMA and CREATE_TABLE privileges on the schema.
# Change the catalog and schema here if necessary.

catalog_name = "main"
schema_name = "default"

source_table_name = "wiki_articles_demo"
source_table_fullname = f"{catalog_name}.{schema_name}.{source_table_name}"

# Uncomment if you want to start from scratch.

# spark.sql(f"DROP TABLE {source_table_fullname}")

source_df = spark.read.parquet("/databricks-datasets/wikipedia-datasets/data-001/en_wikipedia/articles-only-parquet").limit(10)
display(source_df)

Exemple de jeu de données de bloc

La segmentation de l’exemple de jeu de données vous permet d’éviter de dépasser la limite de contexte du modèle d’incorporation. Le modèle GTE prend en charge jusqu’à 8192 jetons. Toutefois, Databricks vous recommande de fractionner les données en blocs de contexte plus petits afin de pouvoir alimenter un plus large éventail d’exemples dans le modèle de raisonnement de votre application RAG.

import tiktoken
import pandas as pd

# The GTE model has been trained on a max context lenth of 8192 tokens.
max_chunk_tokens = 8192
encoding = tiktoken.get_encoding("cl100k_base")

def chunk_text(text):
    # Encode and then decode within the UDF
    tokens = encoding.encode(text)
    chunks = []
    while tokens:
        chunk_tokens = tokens[:max_chunk_tokens]
        chunk_text = encoding.decode(chunk_tokens)
        chunks.append(chunk_text)
        tokens = tokens[max_chunk_tokens:]
    return chunks

# Process the data and store in a new list
pandas_df = source_df.toPandas()
processed_data = []
for index, row in pandas_df.iterrows():
    text_chunks = chunk_text(row['text'])
    chunk_no = 0
    for chunk in text_chunks:
        row_data = row.to_dict()

        # replace the id column with a new unique chunk id
        # and the text column with the text chunk
        row_data['id'] = f"{row['id']}_{chunk_no}"
        row_data['text'] = chunk

        processed_data.append(row_data)
        chunk_no += 1

chunked_pandas_df = pd.DataFrame(processed_data)
chunked_spark_df = spark.createDataFrame(chunked_pandas_df)

# Write the chunked DataFrame to a Delta table
spark.sql(f"DROP TABLE IF EXISTS {source_table_fullname}")
chunked_spark_df.write.format("delta") \
    .option("delta.enableChangeDataFeed", "true") \
    .saveAsTable(source_table_fullname)

display(spark.sql(f"SELECT * FROM {source_table_fullname}"))

Créer un point de terminaison de recherche vectorielle

vector_search_endpoint_name = "vector-search-demo-endpoint"

vsc.create_endpoint(
    name=vector_search_endpoint_name,
    endpoint_type="STANDARD" # or "STORAGE_OPTIMIZED"
)

vsc.get_endpoint(
  name=vector_search_endpoint_name
)

Créer un index vectoriel

# Vector index
vs_index = f"{source_table_name}_gte_index"
vs_index_fullname = f"{catalog_name}.{schema_name}.{vs_index}"

embedding_model_endpoint = "databricks-gte-large-en"

index = vsc.create_delta_sync_index(
  endpoint_name=vector_search_endpoint_name,
  source_table_name=source_table_fullname,
  index_name=vs_index_fullname,
  pipeline_type='TRIGGERED',
  primary_key="id",
  embedding_source_column="text",
  embedding_model_endpoint_name=embedding_model_endpoint
)
index.describe()['status']['message']

# Wait for index to come online. Expect this command to take several minutes.
# You can also track the status of the index build in Catalog Explorer in the
# Overview tab for the vector index.
import time
index = vsc.get_index(endpoint_name=vector_search_endpoint_name,index_name=vs_index_fullname)
while not index.describe().get('status')['ready']:
  print("Waiting for index to be ready...")
  time.sleep(30)
print("Index is ready!")
index.describe()

Recherche de similarité

Les cellules suivantes montrent comment interroger l’index vector pour rechercher des documents similaires.

results = index.similarity_search(
  query_text="Greek myths",
  columns=["id", "text", "title"],
  num_results=5
  )
rows = results['result']['data_array']
for (id, text, title, score) in rows:
  if len(text) > 32:
    # trim text output for readability
    text = text[0:32] + "..."
  print(f"id: {id}  title: {title} text: '{text}' score: {score}")

# Search with a filter. Note that the syntax depends on the endpoint type.

# Standard endpoint syntax
results = index.similarity_search(
  query_text="Greek myths",
  columns=["id", "text", "title"],
  num_results=5,
  filters={"title NOT": "Hercules"}
)

# Storage-optimized endpoint syntax
# results = index.similarity_search(
#   query_text="Greek myths",
#   columns=["id", "text", "title"],
#   num_results=5,
#   filters='title != "Hercules"'
#   )


rows = results['result']['data_array']
for (id, text, title, score) in rows:
  if len(text) > 32:
    # trim text output for readability
    text = text[0:32] + "..."
  print(f"id: {id}  title: {title} text: '{text}' score: {score}")

Supprimer l’index vectoriel

vsc.delete_index(
  endpoint_name=vector_search_endpoint_name,
  index_name=vs_index_fullname
)

Exemple de bloc-notes

Exemple de modèle d'intégration de base (GTE) pour la recherche vectorielle

Obtenir un ordinateur portable