部署批处理推理管道

2025-10-07

重要

此功能目前以公共预览版提供。

本页演示如何将 AI Functions 集成到其他 Databricks 数据和 AI 产品中，以生成完整的批量推理管道。这些管道可以执行端到端工作流，包括引入、预处理、推理和后期处理。管道可以在 SQL 或 Python 中创作，并部署为：

Lakeflow 声明性管道
使用 Databricks 工作流的计划工作流
使用结构化流式处理进行流式推理工作流

要求

在基础模型 API 支持区域中的工作区。
使用 AI Functions 的批处理推理工作负荷需要 Databricks Runtime 15.4 LTS 或更高版本。
查询 Unity Catalog 中 Delta 表的权限，以使用其中包含的数据。
将 pipelines.channel 表中属性设置为“预览”以使用 ai_query()。请参阅要求以获取示例查询。

在 Lakeflow 声明性数据管道上执行增量批次推断

以下示例使用 Lakeflow 声明性管道执行增量批处理推理，以便在数据持续更新时使用。

步骤 1：从存储卷中引入原始新闻数据

SQL

CREATE OR REFRESH STREAMING TABLE news_raw
COMMENT "Raw news articles ingested from volume."
AS SELECT *
FROM STREAM(read_files(
  '/Volumes/databricks_news_summarization_benchmarking_data/v01/csv',
  format => 'csv',
  header => true,
  mode => 'PERMISSIVE',
  multiLine => 'true'
));

Python

导入包并将 LLM 响应的 JSON 架构定义为 Python 变量

from pyspark import pipelines as dp
from pyspark.sql.functions import expr, get_json_object, concat

news_extraction_schema = (
    '{"type": "json_schema", "json_schema": {"name": "news_extraction", '
    '"schema": {"type": "object", "properties": {"title": {"type": "string"}, '
    '"category": {"type": "string", "enum": ["Politics", "Sports", "Technology", '
    '"Health", "Entertainment", "Business"]}}}, "strict": true}}'
)

从 Unity 目录卷引入数据。

@dp.table(
  comment="Raw news articles ingested from volume."
)
def news_raw():
  return (
    spark.readStream
      .format("cloudFiles")
      .option("cloudFiles.format", "csv")
      .option("header", True)
      .option("mode", "PERMISSIVE")
      .option("multiLine", "true")
      .load("/Volumes/databricks_news_summarization_benchmarking_data/v01/csv")
  )

步骤 2：应用 LLM 推理以提取标题和类别

SQL


CREATE OR REFRESH MATERIALIZED VIEW news_categorized
COMMENT "Extract category and title from news articles using LLM inference."
AS
SELECT
  inputs,
  ai_query(
    "databricks-meta-llama-3-3-70b-instruct",
    "Extract the category of the following news article: " || inputs,
    responseFormat => '{
      "type": "json_schema",
      "json_schema": {
        "name": "news_extraction",
        "schema": {
          "type": "object",
          "properties": {
            "title": { "type": "string" },
            "category": {
              "type": "string",
              "enum": ["Politics", "Sports", "Technology", "Health", "Entertainment", "Business"]
            }
          }
        },
        "strict": true
      }
    }'
  ) AS meta_data
FROM news_raw
LIMIT 2;

Python

@dp.materialized_view(
  comment="Extract category and title from news articles using LLM inference."
)
def news_categorized():
  # Limit the number of rows to 2 as in the SQL version
  df_raw = spark.read.table("news_raw").limit(2)
  # Inject the JSON schema variable into the ai_query call using an f-string.
  return df_raw.withColumn(
    "meta_data",
    expr(
      f"ai_query('databricks-meta-llama-3-3-70b-instruct', "
      f"concat('Extract the category of the following news article: ', inputs), "
      f"responseFormat => '{news_extraction_schema}')"
    )
  )

步骤 3：在汇总之前验证 LLM 推理输出

SQL

CREATE OR REFRESH MATERIALIZED VIEW news_validated (
  CONSTRAINT valid_title EXPECT (size(split(get_json_object(meta_data, '$.title'), ' ')) >= 3),
  CONSTRAINT valid_category EXPECT (get_json_object(meta_data, '$.category') IN ('Politics', 'Sports', 'Technology', 'Health', 'Entertainment', 'Business'))
)
COMMENT "Validated news articles ensuring the title has at least 3 words and the category is valid."
AS
SELECT *
FROM news_categorized;

Python

@dp.materialized_view(
  comment="Validated news articles ensuring the title has at least 3 words and the category is valid."
)
@dp.expect("valid_title", "size(split(get_json_object(meta_data, '$.title'), ' ')) >= 3")
@dp.expect_or_fail("valid_category", "get_json_object(meta_data, '$.category') IN ('Politics', 'Sports', 'Technology', 'Health', 'Entertainment', 'Business')")
def news_validated():
  return spark.read.table("news_categorized")

步骤 4：汇总已验证数据中的新闻文章

SQL

CREATE OR REFRESH MATERIALIZED VIEW news_summarized
COMMENT "Summarized political news articles after validation."
AS
SELECT
  get_json_object(meta_data, '$.category') as category,
  get_json_object(meta_data, '$.title') as title,
  ai_query(
    "databricks-meta-llama-3-3-70b-instruct",
    "Summarize the following political news article in 2-3 sentences: " || inputs
  ) AS summary
FROM news_validated;

Python

@dp.materialized_view(
  comment="Summarized political news articles after validation."
)
def news_summarized():
  df = spark.read.table("news_validated")
  return df.select(
    get_json_object("meta_data", "$.category").alias("category"),
    get_json_object("meta_data", "$.title").alias("title"),
    expr(
      "ai_query('databricks-meta-llama-3-3-70b-instruct', "
      "concat('Summarize the following political news article in 2-3 sentences: ', inputs))"
    ).alias("summary")
  )

使用 Databricks 工作流自动执行批处理推理作业

计划批处理推理任务并自动化 AI 管道。

SQL

SELECT
   *,
   ai_query('databricks-meta-llama-3-3-70b-instruct', request => concat("You are an opinion mining service. Given a piece of text, output an array of json results that extracts key user opinions, a classification, and a Positive, Negative, Neutral, or Mixed sentiment about that subject.
AVAILABLE CLASSIFICATIONS
Quality, Service, Design, Safety, Efficiency, Usability, Price
Examples below:
DOCUMENT
I got soup. It really did take only 20 minutes to make some pretty good soup. The noises it makes when it's blending are somewhat terrifying, but it gives a little beep to warn you before it does that. It made three or four large servings of soup. It's a single layer of steel, so the outside gets pretty hot. It can be hard to unplug the lid without knocking the blender against the side, which is not a nice sound. The soup was good and the recipes it comes with look delicious, but I'm not sure I'll use it often. 20 minutes of scary noises from the kitchen when I already need comfort food is not ideal for me. But if you aren't sensitive to loud sounds it does exactly what it says it does..
RESULT
[
 {'Classification': 'Efficiency', 'Comment': 'only 20 minutes','Sentiment': 'Positive'},
 {'Classification': 'Quality','Comment': 'pretty good soup','Sentiment': 'Positive'},
 {'Classification': 'Usability', 'Comment': 'noises it makes when it's blending are somewhat terrifying', 'Sentiment': 'Negative'},
 {'Classification': 'Safety','Comment': 'outside gets pretty hot','Sentiment': 'Negative'},
 {'Classification': 'Design','Comment': 'Hard to unplug the lid without knocking the blender against the side, which is not a nice sound', 'Sentiment': 'Negative'}
]
DOCUMENT
", REVIEW_TEXT, '\n\nRESULT\n')) as result
FROM catalog.schema.product_reviews
LIMIT 10

Python


import json
from pyspark.sql.functions import expr

# Define the opinion mining prompt as a multi-line string.
opinion_prompt = """You are an opinion mining service. Given a piece of text, output an array of json results that extracts key user opinions, a classification, and a Positive, Negative, Neutral, or Mixed sentiment about that subject.
AVAILABLE CLASSIFICATIONS
Quality, Service, Design, Safety, Efficiency, Usability, Price
Examples below:
DOCUMENT
I got soup. It really did take only 20 minutes to make some pretty good soup.The noises it makes when it's blending are somewhat terrifying, but it gives a little beep to warn you before it does that.It made three or four large servings of soup.It's a single layer of steel, so the outside gets pretty hot. It can be hard to unplug the lid without knocking the blender against the side, which is not a nice sound.The soup was good and the recipes it comes with look delicious, but I'm not sure I'll use it often. 20 minutes of scary noises from the kitchen when I already need comfort food is not ideal for me. But if you aren't sensitive to loud sounds it does exactly what it says it does.
RESULT
[
 {'Classification': 'Efficiency', 'Comment': 'only 20 minutes','Sentiment': 'Positive'},
 {'Classification': 'Quality','Comment': 'pretty good soup','Sentiment': 'Positive'},
 {'Classification': 'Usability', 'Comment': 'noises it makes when it's blending are somewhat terrifying', 'Sentiment': 'Negative'},
 {'Classification': 'Safety','Comment': 'outside gets pretty hot','Sentiment': 'Negative'},
 {'Classification': 'Design','Comment': 'Hard to unplug the lid without knocking the blender against the side, which is not a nice sound', 'Sentiment': 'Negative'}
]
DOCUMENT
"""

# Escape the prompt so it can be safely embedded in the SQL expression.
escaped_prompt = json.dumps(opinion_prompt)

# Read the source table and limit to 10 rows.
df = spark.table("catalog.schema.product_reviews").limit(10)

# Apply the LLM inference to each row, concatenating the prompt, the review text, and the tail string.
result_df = df.withColumn(
    "result",
    expr(f"ai_query('databricks-meta-llama-3-3-70b-instruct', request => concat({escaped_prompt}, REVIEW_TEXT, '\\n\\nRESULT\\n'))")
)

# Display the result DataFrame.
display(result_df)

使用结构化流式处理的 AI 函数

使用 ai_query 和结构化流式处理在微批处理或近实时方案中应用 AI 推理。

步骤 1. 读取静态 Delta 表

将静态 Delta 表作为流来读取。


from pyspark.sql import SparkSession
import pyspark.sql.functions as F

spark = SparkSession.builder.getOrCreate()

# Spark processes all existing rows exactly once in the first micro-batch.
df = spark.table("enterprise.docs")  # Replace with your table name containing enterprise documents
df.repartition(50).write.format("delta").mode("overwrite").saveAsTable("enterprise.docs")
df_stream = spark.readStream.format("delta").option("maxBytesPerTrigger", "50K").table("enterprise.docs")

# Define the prompt outside the SQL expression.
prompt = (
    "You are provided with an enterprise document. Summarize the key points in a concise paragraph. "
    "Do not include extra commentary or suggestions. Document: "
)

步骤 2. 应用 `ai_query`

Spark 仅处理静态数据一次，除非新行到达表中。


df_transformed = df_stream.select(
    "document_text",
    F.expr(f"""
      ai_query(
        'databricks-meta-llama-3-1-8b-instruct',
        CONCAT('{prompt}', document_text)
      )
    """).alias("summary")
)

步骤 3：编写汇总输出

将汇总输出写入另一个 Delta 表


# Time-based triggers apply, but only the first trigger processes all existing static data.
query = df_transformed.writeStream \
    .format("delta") \
    .option("checkpointLocation", "/tmp/checkpoints/_docs_summary") \
    .outputMode("append") \
    .toTable("enterprise.docs_summary")

query.awaitTermination()

反馈

此页面是否有帮助？

通过

部署批处理推理管道

要求

在 Lakeflow 声明性数据管道上执行增量批次推断

步骤 1：从存储卷中引入原始新闻数据

SQL

Python

步骤 2：应用 LLM 推理以提取标题和类别

SQL

Python

步骤 3：在汇总之前验证 LLM 推理输出

SQL

Python

步骤 4：汇总已验证数据中的新闻文章

SQL

Python

使用 Databricks 工作流自动执行批处理推理作业

SQL

Python

使用结构化流式处理的 AI 函数

步骤 1. 读取静态 Delta 表

步骤 2. 应用 ai_query

步骤 3：编写汇总输出

反馈

其他资源

步骤 2. 应用 `ai_query`