Exemple de modèle d'intégration fondamental AI Search (GTE)
Ce Notebook montre comment utiliser le SDK Python AI Search, qui fournit AISearchClient comme API principale pour travailler avec AI Search.
Ce Notebook utilise les APIs de modèle de fondation Databricks pour accéder au modèle d'embeddings GTE afin de générer des embeddings.
Python
%pip install --upgrade --force-reinstall databricks-ai-search
dbutils.library.restartPython()
Python
from databricks.ai_search.client import AISearchClient
vsc = AISearchClient(disable_notice=True)
Python
help(AISearchClient)
Chargez un dataset de démonstration dans la table Delta source.
Ce qui suit crée la table Delta source.
Python
# 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"
Python
source_table_name = "wiki_articles_demo"
source_table_fullname = f"{catalog_name}.{schema_name}.{source_table_name}"
Python
# Uncomment if you want to start from scratch.
# spark.sql(f"DROP TABLE {source_table_fullname}")
Python
source_df = spark.read.parquet("/databricks-datasets/wikipedia-datasets/data-001/en_wikipedia/articles-only-parquet").limit(10)
display(source_df)
Extrait de dataset
Le découpage du dataset d’échantillons vous aide à éviter de dépasser la limite de contexte du modèle d’intégration. Le modèle GTE prend en charge jusqu’à 8 192 tokens. Cependant, Databricks vous recommande de diviser les données en fragments de contexte plus petits afin que vous puissiez alimenter un plus grand nombre d’exemples dans le modèle de raisonnement de votre application RAG.
Python
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)
Python
display(spark.sql(f"SELECT * FROM {source_table_fullname}"))
Créer un Endpoint
Python
ai_search_endpoint_name = "ai-search-demo-endpoint"
Python
vsc.create_endpoint(
name=ai_search_endpoint_name,
endpoint_type="STANDARD" # or "STORAGE_OPTIMIZED"
)
Python
vsc.get_endpoint(
name=ai_search_endpoint_name
)
Créer un index
Python
# AI Search index
vs_index = f"{source_table_name}_gte_index"
vs_index_fullname = f"{catalog_name}.{schema_name}.{vs_index}"
embedding_model_endpoint = "databricks-qwen3-embedding-0-6b"
Python
index = vsc.create_delta_sync_index(
endpoint_name=ai_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']
Python
# 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 index.
import time
index = vsc.get_index(endpoint_name=ai_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 pour trouver des documents similaires.
Python
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}")
Python
# 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
Python
vsc.delete_index(
endpoint_name=ai_search_endpoint_name,
index_name=vs_index_fullname
)