高级文本转 SQL 工作流¶

本指南将展示如何通过我们的工作流语法为您的数据设置文本转 SQL 工作流。

这种方案让您能够灵活地运用附加技术增强文本转 SQL 功能。我们将在以下章节展示这些技术：

1. 查询时表检索：在文本转 SQL 提示中动态检索相关表
2. 查询时样本行检索：对每行数据进行嵌入/索引，并在文本转 SQL 提示中为每个表动态检索示例行

我们的开箱即用工作流包含 NLSQLTableQueryEngine 和 SQLTableRetrieverQueryEngine（若想查看使用这些模块的文本转 SQL 指南，请访问此处）。本指南实现了这些模块的高级版本，为您提供最大灵活性以适应个性化场景。

注意： 任何文本转 SQL 应用都应注意，执行任意 SQL 查询可能存在安全风险。建议根据需要采取预防措施，例如使用受限角色、只读数据库、沙盒环境等。

数据加载与导入¶

加载数据¶

我们采用WikiTableQuestions数据集（Pasupat和Liang 2015）作为测试数据集。

我们会遍历指定文件夹中的所有csv文件，将每个文件存储到sqlite数据库中（后续将在每个表结构上构建对象索引）。

In [ ]:

%pip install llama-index-llms-openai

%pip install llama-index-llms-openai

In [ ]:

!wget "https://github.com/ppasupat/WikiTableQuestions/releases/download/v1.0.2/WikiTableQuestions-1.0.2-compact.zip" -O data.zip
!unzip data.zip

!wget "https://github.com/ppasupat/WikiTableQuestions/releases/download/v1.0.2/WikiTableQuestions-1.0.2-compact.zip" -O data.zip !unzip data.zip

In [ ]:

import pandas as pd
from pathlib import Path

data_dir = Path("./WikiTableQuestions/csv/200-csv")
csv_files = sorted([f for f in data_dir.glob("*.csv")])
dfs = []
for csv_file in csv_files:
    print(f"processing file: {csv_file}")
    try:
        df = pd.read_csv(csv_file)
        dfs.append(df)
    except Exception as e:
        print(f"Error parsing {csv_file}: {str(e)}")

import pandas as pd from pathlib import Path data_dir = Path("./WikiTableQuestions/csv/200-csv") csv_files = sorted([f for f in data_dir.glob("*.csv")]) dfs = [] for csv_file in csv_files: print(f"processing file: {csv_file}") try: df = pd.read_csv(csv_file) dfs.append(df) except Exception as e: print(f"Error parsing {csv_file}: {str(e)}")

从每个表格中提取表名与摘要¶

此处我们使用 gpt-4o-mini 模型，通过 Pydantic 程序从每个表格中提取带下划线的表名及其摘要。

In [ ]:

tableinfo_dir = "WikiTableQuestions_TableInfo"
!mkdir {tableinfo_dir}

tableinfo_dir = "WikiTableQuestions_TableInfo" !mkdir {tableinfo_dir}

mkdir: WikiTableQuestions_TableInfo: File exists

In [ ]:

from llama_index.core.prompts import ChatPromptTemplate
from llama_index.core.bridge.pydantic import BaseModel, Field
from llama_index.llms.openai import OpenAI
from llama_index.core.llms import ChatMessage


class TableInfo(BaseModel):
    """Information regarding a structured table."""

    table_name: str = Field(
        ..., description="table name (must be underscores and NO spaces)"
    )
    table_summary: str = Field(
        ..., description="short, concise summary/caption of the table"
    )


prompt_str = """\
Give me a summary of the table with the following JSON format.

- The table name must be unique to the table and describe it while being concise. 
- Do NOT output a generic table name (e.g. table, my_table).

Do NOT make the table name one of the following: {exclude_table_name_list}

Table:
{table_str}

Summary: """
prompt_tmpl = ChatPromptTemplate(
    message_templates=[ChatMessage.from_str(prompt_str, role="user")]
)

llm = OpenAI(model="gpt-4o-mini")

from llama_index.core.prompts import ChatPromptTemplate from llama_index.core.bridge.pydantic import BaseModel, Field from llama_index.llms.openai import OpenAI from llama_index.core.llms import ChatMessage class TableInfo(BaseModel): """Information regarding a structured table.""" table_name: str = Field( ..., description="table name (must be underscores and NO spaces)" ) table_summary: str = Field( ..., description="short, concise summary/caption of the table" ) prompt_str = """\ Give me a summary of the table with the following JSON format. - The table name must be unique to the table and describe it while being concise. - Do NOT output a generic table name (e.g. table, my_table). Do NOT make the table name one of the following: {exclude_table_name_list} Table: {table_str} Summary: """ prompt_tmpl = ChatPromptTemplate( message_templates=[ChatMessage.from_str(prompt_str, role="user")] ) llm = OpenAI(model="gpt-4o-mini")

In [ ]:

import json


def _get_tableinfo_with_index(idx: int) -> str:
    results_gen = Path(tableinfo_dir).glob(f"{idx}_*")
    results_list = list(results_gen)
    if len(results_list) == 0:
        return None
    elif len(results_list) == 1:
        path = results_list[0]
        return TableInfo.parse_file(path)
    else:
        raise ValueError(
            f"More than one file matching index: {list(results_gen)}"
        )


table_names = set()
table_infos = []
for idx, df in enumerate(dfs):
    table_info = _get_tableinfo_with_index(idx)
    if table_info:
        table_infos.append(table_info)
    else:
        while True:
            df_str = df.head(10).to_csv()
            table_info = llm.structured_predict(
                TableInfo,
                prompt_tmpl,
                table_str=df_str,
                exclude_table_name_list=str(list(table_names)),
            )
            table_name = table_info.table_name
            print(f"Processed table: {table_name}")
            if table_name not in table_names:
                table_names.add(table_name)
                break
            else:
                # try again
                print(f"Table name {table_name} already exists, trying again.")
                pass

        out_file = f"{tableinfo_dir}/{idx}_{table_name}.json"
        json.dump(table_info.dict(), open(out_file, "w"))
    table_infos.append(table_info)

import json def _get_tableinfo_with_index(idx: int) -> str: results_gen = Path(tableinfo_dir).glob(f"{idx}_*") results_list = list(results_gen) if len(results_list) == 0: return None elif len(results_list) == 1: path = results_list[0] return TableInfo.parse_file(path) else: raise ValueError( f"More than one file matching index: {list(results_gen)}" ) table_names = set() table_infos = [] for idx, df in enumerate(dfs): table_info = _get_tableinfo_with_index(idx) if table_info: table_infos.append(table_info) else: while True: df_str = df.head(10).to_csv() table_info = llm.structured_predict( TableInfo, prompt_tmpl, table_str=df_str, exclude_table_name_list=str(list(table_names)), ) table_name = table_info.table_name print(f"Processed table: {table_name}") if table_name not in table_names: table_names.add(table_name) break else: # try again print(f"Table name {table_name} already exists, trying again.") pass out_file = f"{tableinfo_dir}/{idx}_{table_name}.json" json.dump(table_info.dict(), open(out_file, "w")) table_infos.append(table_info)

将数据存入 SQL 数据库¶

我们使用流行的 SQL 数据库工具包 sqlalchemy 来加载所有数据表。

In [ ]:

# put data into sqlite db
from sqlalchemy import (
    create_engine,
    MetaData,
    Table,
    Column,
    String,
    Integer,
)
import re


# Function to create a sanitized column name
def sanitize_column_name(col_name):
    # Remove special characters and replace spaces with underscores
    return re.sub(r"\W+", "_", col_name)


# Function to create a table from a DataFrame using SQLAlchemy
def create_table_from_dataframe(
    df: pd.DataFrame, table_name: str, engine, metadata_obj
):
    # Sanitize column names
    sanitized_columns = {col: sanitize_column_name(col) for col in df.columns}
    df = df.rename(columns=sanitized_columns)

    # Dynamically create columns based on DataFrame columns and data types
    columns = [
        Column(col, String if dtype == "object" else Integer)
        for col, dtype in zip(df.columns, df.dtypes)
    ]

    # Create a table with the defined columns
    table = Table(table_name, metadata_obj, *columns)

    # Create the table in the database
    metadata_obj.create_all(engine)

    # Insert data from DataFrame into the table
    with engine.connect() as conn:
        for _, row in df.iterrows():
            insert_stmt = table.insert().values(**row.to_dict())
            conn.execute(insert_stmt)
        conn.commit()


# engine = create_engine("sqlite:///:memory:")
engine = create_engine("sqlite:///wiki_table_questions.db")
metadata_obj = MetaData()
for idx, df in enumerate(dfs):
    tableinfo = _get_tableinfo_with_index(idx)
    print(f"Creating table: {tableinfo.table_name}")
    create_table_from_dataframe(df, tableinfo.table_name, engine, metadata_obj)

# put data into sqlite db from sqlalchemy import ( create_engine, MetaData, Table, Column, String, Integer, ) import re # Function to create a sanitized column name def sanitize_column_name(col_name): # Remove special characters and replace spaces with underscores return re.sub(r"\W+", "_", col_name) # Function to create a table from a DataFrame using SQLAlchemy def create_table_from_dataframe( df: pd.DataFrame, table_name: str, engine, metadata_obj ): # Sanitize column names sanitized_columns = {col: sanitize_column_name(col) for col in df.columns} df = df.rename(columns=sanitized_columns) # Dynamically create columns based on DataFrame columns and data types columns = [ Column(col, String if dtype == "object" else Integer) for col, dtype in zip(df.columns, df.dtypes) ] # Create a table with the defined columns table = Table(table_name, metadata_obj, *columns) # Create the table in the database metadata_obj.create_all(engine) # Insert data from DataFrame into the table with engine.connect() as conn: for _, row in df.iterrows(): insert_stmt = table.insert().values(**row.to_dict()) conn.execute(insert_stmt) conn.commit() # engine = create_engine("sqlite:///:memory:") engine = create_engine("sqlite:///wiki_table_questions.db") metadata_obj = MetaData() for idx, df in enumerate(dfs): tableinfo = _get_tableinfo_with_index(idx) print(f"Creating table: {tableinfo.table_name}") create_table_from_dataframe(df, tableinfo.table_name, engine, metadata_obj)

In [ ]:

# # setup Arize Phoenix for logging/observability
# import phoenix as px
# import llama_index.core

# px.launch_app()
# llama_index.core.set_global_handler("arize_phoenix")

# # setup Arize Phoenix for logging/observability # import phoenix as px # import llama_index.core # px.launch_app() # llama_index.core.set_global_handler("arize_phoenix")

高级能力 1：支持查询时表检索的文本转 SQL¶

现在我们将展示如何搭建一个端到端的文本转 SQL 系统，并实现表检索功能。

定义核心模块¶

此处我们定义核心功能模块：

1. 对象索引 + 检索器（用于存储表结构）
2. SQLDatabase 对象（连接上述表） + SQLRetriever
3. 文本转 SQL 提示模板
4. 响应合成提示模板
5. 大语言模型（LLM）

对象索引、检索器、SQL数据库

In [ ]:

from llama_index.core.objects import (
    SQLTableNodeMapping,
    ObjectIndex,
    SQLTableSchema,
)
from llama_index.core import SQLDatabase, VectorStoreIndex

sql_database = SQLDatabase(engine)

table_node_mapping = SQLTableNodeMapping(sql_database)
table_schema_objs = [
    SQLTableSchema(table_name=t.table_name, context_str=t.table_summary)
    for t in table_infos
]  # add a SQLTableSchema for each table

obj_index = ObjectIndex.from_objects(
    table_schema_objs,
    table_node_mapping,
    VectorStoreIndex,
)
obj_retriever = obj_index.as_retriever(similarity_top_k=3)

from llama_index.core.objects import ( SQLTableNodeMapping, ObjectIndex, SQLTableSchema, ) from llama_index.core import SQLDatabase, VectorStoreIndex sql_database = SQLDatabase(engine) table_node_mapping = SQLTableNodeMapping(sql_database) table_schema_objs = [ SQLTableSchema(table_name=t.table_name, context_str=t.table_summary) for t in table_infos ] # add a SQLTableSchema for each table obj_index = ObjectIndex.from_objects( table_schema_objs, table_node_mapping, VectorStoreIndex, ) obj_retriever = obj_index.as_retriever(similarity_top_k=3)

SQLRetriever + 表格解析器

In [ ]:

from llama_index.core.retrievers import SQLRetriever
from typing import List

sql_retriever = SQLRetriever(sql_database)


def get_table_context_str(table_schema_objs: List[SQLTableSchema]):
    """Get table context string."""
    context_strs = []
    for table_schema_obj in table_schema_objs:
        table_info = sql_database.get_single_table_info(
            table_schema_obj.table_name
        )
        if table_schema_obj.context_str:
            table_opt_context = " The table description is: "
            table_opt_context += table_schema_obj.context_str
            table_info += table_opt_context

        context_strs.append(table_info)
    return "\n\n".join(context_strs)

from llama_index.core.retrievers import SQLRetriever from typing import List sql_retriever = SQLRetriever(sql_database) def get_table_context_str(table_schema_objs: List[SQLTableSchema]): """Get table context string.""" context_strs = [] for table_schema_obj in table_schema_objs: table_info = sql_database.get_single_table_info( table_schema_obj.table_name ) if table_schema_obj.context_str: table_opt_context = " The table description is: " table_opt_context += table_schema_obj.context_str table_info += table_opt_context context_strs.append(table_info) return "\n\n".join(context_strs)

文本转SQL提示词 + 输出解析器

In [ ]:

from llama_index.core.prompts.default_prompts import DEFAULT_TEXT_TO_SQL_PROMPT
from llama_index.core import PromptTemplate
from llama_index.core.llms import ChatResponse


def parse_response_to_sql(chat_response: ChatResponse) -> str:
    """Parse response to SQL."""
    response = chat_response.message.content
    sql_query_start = response.find("SQLQuery:")
    if sql_query_start != -1:
        response = response[sql_query_start:]
        # TODO: move to removeprefix after Python 3.9+
        if response.startswith("SQLQuery:"):
            response = response[len("SQLQuery:") :]
    sql_result_start = response.find("SQLResult:")
    if sql_result_start != -1:
        response = response[:sql_result_start]
    return response.strip().strip("```").strip()


text2sql_prompt = DEFAULT_TEXT_TO_SQL_PROMPT.partial_format(
    dialect=engine.dialect.name
)
print(text2sql_prompt.template)

from llama_index.core.prompts.default_prompts import DEFAULT_TEXT_TO_SQL_PROMPT from llama_index.core import PromptTemplate from llama_index.core.llms import ChatResponse def parse_response_to_sql(chat_response: ChatResponse) -> str: """Parse response to SQL.""" response = chat_response.message.content sql_query_start = response.find("SQLQuery:") if sql_query_start != -1: response = response[sql_query_start:] # TODO: move to removeprefix after Python 3.9+ if response.startswith("SQLQuery:"): response = response[len("SQLQuery:") :] sql_result_start = response.find("SQLResult:") if sql_result_start != -1: response = response[:sql_result_start] return response.strip().strip("```").strip() text2sql_prompt = DEFAULT_TEXT_TO_SQL_PROMPT.partial_format( dialect=engine.dialect.name ) print(text2sql_prompt.template)

Given an input question, first create a syntactically correct {dialect} query to run, then look at the results of the query and return the answer. You can order the results by a relevant column to return the most interesting examples in the database.

Never query for all the columns from a specific table, only ask for a few relevant columns given the question.

Pay attention to use only the column names that you can see in the schema description. Be careful to not query for columns that do not exist. Pay attention to which column is in which table. Also, qualify column names with the table name when needed. You are required to use the following format, each taking one line:

Question: Question here
SQLQuery: SQL Query to run
SQLResult: Result of the SQLQuery
Answer: Final answer here

Only use tables listed below.
{schema}

Question: {query_str}
SQLQuery: 

响应合成提示

In [ ]:

response_synthesis_prompt_str = (
    "Given an input question, synthesize a response from the query results.\n"
    "Query: {query_str}\n"
    "SQL: {sql_query}\n"
    "SQL Response: {context_str}\n"
    "Response: "
)
response_synthesis_prompt = PromptTemplate(
    response_synthesis_prompt_str,
)

response_synthesis_prompt_str = ( "Given an input question, synthesize a response from the query results.\n" "Query: {query_str}\n" "SQL: {sql_query}\n" "SQL Response: {context_str}\n" "Response: " ) response_synthesis_prompt = PromptTemplate( response_synthesis_prompt_str, )

In [ ]:

# llm = OpenAI(model="gpt-3.5-turbo")
llm = OpenAI(model="gpt-4o-mini")

# llm = OpenAI(model="gpt-3.5-turbo") llm = OpenAI(model="gpt-4o-mini")

定义工作流¶

现在各组件已就绪，让我们来定义完整的工作流程！

In [ ]:

from llama_index.core.workflow import (
    Workflow,
    StartEvent,
    StopEvent,
    step,
    Context,
    Event,
)


class TableRetrieveEvent(Event):
    """Result of running table retrieval."""

    table_context_str: str
    query: str


class TextToSQLEvent(Event):
    """Text-to-SQL event."""

    sql: str
    query: str


class TextToSQLWorkflow1(Workflow):
    """Text-to-SQL Workflow that does query-time table retrieval."""

    def __init__(
        self,
        obj_retriever,
        text2sql_prompt,
        sql_retriever,
        response_synthesis_prompt,
        llm,
        *args,
        **kwargs
    ) -> None:
        """Init params."""
        super().__init__(*args, **kwargs)
        self.obj_retriever = obj_retriever
        self.text2sql_prompt = text2sql_prompt
        self.sql_retriever = sql_retriever
        self.response_synthesis_prompt = response_synthesis_prompt
        self.llm = llm

    @step
    def retrieve_tables(
        self, ctx: Context, ev: StartEvent
    ) -> TableRetrieveEvent:
        """Retrieve tables."""
        table_schema_objs = self.obj_retriever.retrieve(ev.query)
        table_context_str = get_table_context_str(table_schema_objs)
        return TableRetrieveEvent(
            table_context_str=table_context_str, query=ev.query
        )

    @step
    def generate_sql(
        self, ctx: Context, ev: TableRetrieveEvent
    ) -> TextToSQLEvent:
        """Generate SQL statement."""
        fmt_messages = self.text2sql_prompt.format_messages(
            query_str=ev.query, schema=ev.table_context_str
        )
        chat_response = self.llm.chat(fmt_messages)
        sql = parse_response_to_sql(chat_response)
        return TextToSQLEvent(sql=sql, query=ev.query)

    @step
    def generate_response(self, ctx: Context, ev: TextToSQLEvent) -> StopEvent:
        """Run SQL retrieval and generate response."""
        retrieved_rows = self.sql_retriever.retrieve(ev.sql)
        fmt_messages = self.response_synthesis_prompt.format_messages(
            sql_query=ev.sql,
            context_str=str(retrieved_rows),
            query_str=ev.query,
        )
        chat_response = llm.chat(fmt_messages)
        return StopEvent(result=chat_response)

from llama_index.core.workflow import ( Workflow, StartEvent, StopEvent, step, Context, Event, ) class TableRetrieveEvent(Event): """Result of running table retrieval.""" table_context_str: str query: str class TextToSQLEvent(Event): """Text-to-SQL event.""" sql: str query: str class TextToSQLWorkflow1(Workflow): """Text-to-SQL Workflow that does query-time table retrieval.""" def __init__( self, obj_retriever, text2sql_prompt, sql_retriever, response_synthesis_prompt, llm, *args, **kwargs ) -> None: """Init params.""" super().__init__(*args, **kwargs) self.obj_retriever = obj_retriever self.text2sql_prompt = text2sql_prompt self.sql_retriever = sql_retriever self.response_synthesis_prompt = response_synthesis_prompt self.llm = llm @step def retrieve_tables( self, ctx: Context, ev: StartEvent ) -> TableRetrieveEvent: """Retrieve tables.""" table_schema_objs = self.obj_retriever.retrieve(ev.query) table_context_str = get_table_context_str(table_schema_objs) return TableRetrieveEvent( table_context_str=table_context_str, query=ev.query ) @step def generate_sql( self, ctx: Context, ev: TableRetrieveEvent ) -> TextToSQLEvent: """Generate SQL statement.""" fmt_messages = self.text2sql_prompt.format_messages( query_str=ev.query, schema=ev.table_context_str ) chat_response = self.llm.chat(fmt_messages) sql = parse_response_to_sql(chat_response) return TextToSQLEvent(sql=sql, query=ev.query) @step def generate_response(self, ctx: Context, ev: TextToSQLEvent) -> StopEvent: """Run SQL retrieval and generate response.""" retrieved_rows = self.sql_retriever.retrieve(ev.sql) fmt_messages = self.response_synthesis_prompt.format_messages( sql_query=ev.sql, context_str=str(retrieved_rows), query_str=ev.query, ) chat_response = llm.chat(fmt_messages) return StopEvent(result=chat_response)

可视化工作流¶

工作流的一个绝佳特性在于，你既能查看执行流程图，也能追踪最近一次执行的轨迹。

In [ ]:

from llama_index.utils.workflow import draw_all_possible_flows

draw_all_possible_flows(
    TextToSQLWorkflow1, filename="text_to_sql_table_retrieval.html"
)

from llama_index.utils.workflow import draw_all_possible_flows draw_all_possible_flows( TextToSQLWorkflow1, filename="text_to_sql_table_retrieval.html" )

text_to_sql_table_retrieval.html

In [ ]:

from IPython.display import display, HTML

# Read the contents of the HTML file
with open("text_to_sql_table_retrieval.html", "r") as file:
    html_content = file.read()

# Display the HTML content
display(HTML(html_content))

from IPython.display import display, HTML # Read the contents of the HTML file with open("text_to_sql_table_retrieval.html", "r") as file: html_content = file.read() # Display the HTML content display(HTML(html_content))

执行查询操作！¶

现在我们已经准备好对整个工作流执行一些查询了。

In [ ]:

workflow = TextToSQLWorkflow1(
    obj_retriever,
    text2sql_prompt,
    sql_retriever,
    response_synthesis_prompt,
    llm,
    verbose=True,
)

workflow = TextToSQLWorkflow1( obj_retriever, text2sql_prompt, sql_retriever, response_synthesis_prompt, llm, verbose=True, )

In [ ]:

response = await workflow.run(
    query="What was the year that The Notorious B.I.G was signed to Bad Boy?"
)
print(str(response))

response = await workflow.run( query="What was the year that The Notorious B.I.G was signed to Bad Boy?" ) print(str(response))

Running step retrieve_tables
Step retrieve_tables produced event TableRetrieveEvent
Running step generate_sql
Step generate_sql produced event TextToSQLEvent
Running step generate_response
Step generate_response produced event StopEvent
assistant: The Notorious B.I.G was signed to Bad Boy Records in 1993.
VERBOSE: True
> Table Info: Table 'bad_boy_artists_album_release_summary' has columns: Act (VARCHAR), Year_signed (INTEGER), _Albums_released_under_Bad_Boy (VARCHAR), . The table description is: A summary of artists signed to Bad Boy Records along with the year they were signed and the number of albums they released.
Here are some relevant example rows (values in the same order as columns above)
('The Notorious B.I.G', 1993, '5')

> Table Info: Table 'filmography_of_diane_drummond' has columns: Year (INTEGER), Title (VARCHAR), Role (VARCHAR), Notes (VARCHAR), . The table description is: A list of film and television roles played by Diane Drummond from 1995 to 2001.
Here are some relevant example rows (values in the same order as columns above)
(2013, 'L.A. Slasher', 'The Actress', None)

> Table Info: Table 'progressive_rock_album_chart_positions' has columns: Year (INTEGER), Title (VARCHAR), Chart_Positions_UK (VARCHAR), Chart_Positions_US (VARCHAR), Chart_Positions_NL (VARCHAR), Comments (VARCHAR), . The table description is: Chart positions of progressive rock albums in the UK, US, and NL from 1969 to 1981.
Here are some relevant example rows (values in the same order as columns above)
(1977, 'Novella', '–', '46', '–', '1977 (January in US, August in UK, as the band moved to the Warner Bros Music Group)')

VERBOSE: True
> Table Info: Table 'bad_boy_artists_album_release_summary' has columns: Act (VARCHAR), Year_signed (INTEGER), _Albums_released_under_Bad_Boy (VARCHAR), . The table description is: A summary of artists signed to Bad Boy Records along with the year they were signed and the number of albums they released.
Here are some relevant example rows (values in the same order as columns above)
('The Notorious B.I.G', 1993, '5')

> Table Info: Table 'filmography_of_diane_drummond' has columns: Year (INTEGER), Title (VARCHAR), Role (VARCHAR), Notes (VARCHAR), . The table description is: A list of film and television roles played by Diane Drummond from 1995 to 2001.
Here are some relevant example rows (values in the same order as columns above)
(2013, 'L.A. Slasher', 'The Actress', None)

> Table Info: Table 'progressive_rock_album_chart_positions' has columns: Year (INTEGER), Title (VARCHAR), Chart_Positions_UK (VARCHAR), Chart_Positions_US (VARCHAR), Chart_Positions_NL (VARCHAR), Comments (VARCHAR), . The table description is: Chart positions of progressive rock albums in the UK, US, and NL from 1969 to 1981.
Here are some relevant example rows (values in the same order as columns above)
(1977, 'Novella', '–', '46', '–', '1977 (January in US, August in UK, as the band moved to the Warner Bros Music Group)')

In [ ]:

response = await workflow.run(
    query="Who won best director in the 1972 academy awards"
)
print(str(response))

response = await workflow.run( query="Who won best director in the 1972 academy awards" ) print(str(response))

Running step retrieve_tables
Step retrieve_tables produced event TableRetrieveEvent
Running step generate_sql
Step generate_sql produced event TextToSQLEvent
Running step generate_response
Step generate_response produced event StopEvent
assistant: William Friedkin won the Best Director award at the 1972 Academy Awards.

In [ ]:

response = await workflow.run(query="What was the term of Pasquale Preziosa?")
print(str(response))

response = await workflow.run(query="What was the term of Pasquale Preziosa?") print(str(response))

Running step retrieve_tables
Step retrieve_tables produced event TableRetrieveEvent
Running step generate_sql
Step generate_sql produced event TextToSQLEvent
Running step generate_response
Step generate_response produced event StopEvent
assistant: Pasquale Preziosa has been serving since 25 February 2013 and is currently in office as the incumbent.

2. 高级能力 2：支持查询时行检索（及表检索）的文本转 SQL 功能¶

前例中存在一个问题：若用户查询"The Notorious BIG"但数据库中存储的是"The Notorious B.I.G"，生成的SELECT语句很可能无法返回匹配结果。

我们可以通过获取每张表的少量示例行来缓解此问题。简单方案是直接获取前k行，但我们采用了更智能的方式——根据用户查询对行数据进行嵌入、索引和检索，从而为文本转SQL的大语言模型提供最具上下文相关性的信息来生成SQL。

现在我们将扩展工作流程。

为每个表建立索引¶

我们对每个表的行数据进行嵌入/索引处理，最终为每个表生成一个独立的索引。

In [ ]:

from llama_index.core import VectorStoreIndex, load_index_from_storage
from sqlalchemy import text
from llama_index.core.schema import TextNode
from llama_index.core import StorageContext
import os
from pathlib import Path
from typing import Dict


def index_all_tables(
    sql_database: SQLDatabase, table_index_dir: str = "table_index_dir"
) -> Dict[str, VectorStoreIndex]:
    """Index all tables."""
    if not Path(table_index_dir).exists():
        os.makedirs(table_index_dir)

    vector_index_dict = {}
    engine = sql_database.engine
    for table_name in sql_database.get_usable_table_names():
        print(f"Indexing rows in table: {table_name}")
        if not os.path.exists(f"{table_index_dir}/{table_name}"):
            # get all rows from table
            with engine.connect() as conn:
                cursor = conn.execute(text(f'SELECT * FROM "{table_name}"'))
                result = cursor.fetchall()
                row_tups = []
                for row in result:
                    row_tups.append(tuple(row))

            # index each row, put into vector store index
            nodes = [TextNode(text=str(t)) for t in row_tups]

            # put into vector store index (use OpenAIEmbeddings by default)
            index = VectorStoreIndex(nodes)

            # save index
            index.set_index_id("vector_index")
            index.storage_context.persist(f"{table_index_dir}/{table_name}")
        else:
            # rebuild storage context
            storage_context = StorageContext.from_defaults(
                persist_dir=f"{table_index_dir}/{table_name}"
            )
            # load index
            index = load_index_from_storage(
                storage_context, index_id="vector_index"
            )
        vector_index_dict[table_name] = index

    return vector_index_dict


vector_index_dict = index_all_tables(sql_database)

from llama_index.core import VectorStoreIndex, load_index_from_storage from sqlalchemy import text from llama_index.core.schema import TextNode from llama_index.core import StorageContext import os from pathlib import Path from typing import Dict def index_all_tables( sql_database: SQLDatabase, table_index_dir: str = "table_index_dir" ) -> Dict[str, VectorStoreIndex]: """Index all tables.""" if not Path(table_index_dir).exists(): os.makedirs(table_index_dir) vector_index_dict = {} engine = sql_database.engine for table_name in sql_database.get_usable_table_names(): print(f"Indexing rows in table: {table_name}") if not os.path.exists(f"{table_index_dir}/{table_name}"): # get all rows from table with engine.connect() as conn: cursor = conn.execute(text(f'SELECT * FROM "{table_name}"')) result = cursor.fetchall() row_tups = [] for row in result: row_tups.append(tuple(row)) # index each row, put into vector store index nodes = [TextNode(text=str(t)) for t in row_tups] # put into vector store index (use OpenAIEmbeddings by default) index = VectorStoreIndex(nodes) # save index index.set_index_id("vector_index") index.storage_context.persist(f"{table_index_dir}/{table_name}") else: # rebuild storage context storage_context = StorageContext.from_defaults( persist_dir=f"{table_index_dir}/{table_name}" ) # load index index = load_index_from_storage( storage_context, index_id="vector_index" ) vector_index_dict[table_name] = index return vector_index_dict vector_index_dict = index_all_tables(sql_database)

Indexing rows in table: academy_awards_and_nominations_1972
Indexing rows in table: annual_traffic_accident_deaths
Indexing rows in table: bad_boy_artists_album_release_summary
Indexing rows in table: bbc_radio_services_cost_comparison_2012_2013
Indexing rows in table: binary_encoding_probabilities
Indexing rows in table: boxing_match_results_summary
Indexing rows in table: cancer_related_genes_and_functions
Indexing rows in table: diane_drummond_awards_nominations
Indexing rows in table: diane_drummond_oscar_nominations_and_wins
Indexing rows in table: diane_drummond_single_chart_performance
Indexing rows in table: euro_2020_group_stage_results
Indexing rows in table: experiment_drop_events_timeline
Indexing rows in table: filmography_of_diane_drummond
Indexing rows in table: grammy_awards_summary_for_wilco
Indexing rows in table: historical_college_football_records
Indexing rows in table: italian_ministers_term_dates
Indexing rows in table: kodachrome_film_types_and_dates
Indexing rows in table: missing_persons_case_summary
Indexing rows in table: monthly_climate_statistics
Indexing rows in table: monthly_climate_statistics_summary
Indexing rows in table: monthly_weather_statistics
Indexing rows in table: multilingual_greetings_and_phrases
Indexing rows in table: municipalities_merger_summary
Indexing rows in table: new_mexico_government_officials
Indexing rows in table: norwegian_club_performance_summary
Indexing rows in table: ohio_private_schools_summary
Indexing rows in table: progressive_rock_album_chart_positions
Indexing rows in table: regional_airports_usage_summary
Indexing rows in table: south_dakota_radio_stations
Indexing rows in table: triple_crown_winners_summary
Indexing rows in table: uk_ministers_and_titles_history
Indexing rows in table: voter_registration_status_by_party
Indexing rows in table: voter_registration_summary_by_party
Indexing rows in table: yamato_district_population_density

定义扩展表格解析功能¶

我们增强了表格解析的能力，使其不仅能返回相关的表格结构，还能为每个表格结构返回相关的数据行。

现在该功能不仅接收 table_schema_objs（表格检索器的输出），还接收原始 query_str 参数，后者将用于通过向量检索获取相关数据行。

In [ ]:

from llama_index.core.retrievers import SQLRetriever
from typing import List

sql_retriever = SQLRetriever(sql_database)


def get_table_context_and_rows_str(
    query_str: str,
    table_schema_objs: List[SQLTableSchema],
    verbose: bool = False,
):
    """Get table context string."""
    context_strs = []
    for table_schema_obj in table_schema_objs:
        # first append table info + additional context
        table_info = sql_database.get_single_table_info(
            table_schema_obj.table_name
        )
        if table_schema_obj.context_str:
            table_opt_context = " The table description is: "
            table_opt_context += table_schema_obj.context_str
            table_info += table_opt_context

        # also lookup vector index to return relevant table rows
        vector_retriever = vector_index_dict[
            table_schema_obj.table_name
        ].as_retriever(similarity_top_k=2)
        relevant_nodes = vector_retriever.retrieve(query_str)
        if len(relevant_nodes) > 0:
            table_row_context = "\nHere are some relevant example rows (values in the same order as columns above)\n"
            for node in relevant_nodes:
                table_row_context += str(node.get_content()) + "\n"
            table_info += table_row_context

        if verbose:
            print(f"> Table Info: {table_info}")

        context_strs.append(table_info)
    return "\n\n".join(context_strs)

from llama_index.core.retrievers import SQLRetriever from typing import List sql_retriever = SQLRetriever(sql_database) def get_table_context_and_rows_str( query_str: str, table_schema_objs: List[SQLTableSchema], verbose: bool = False, ): """Get table context string.""" context_strs = [] for table_schema_obj in table_schema_objs: # first append table info + additional context table_info = sql_database.get_single_table_info( table_schema_obj.table_name ) if table_schema_obj.context_str: table_opt_context = " The table description is: " table_opt_context += table_schema_obj.context_str table_info += table_opt_context # also lookup vector index to return relevant table rows vector_retriever = vector_index_dict[ table_schema_obj.table_name ].as_retriever(similarity_top_k=2) relevant_nodes = vector_retriever.retrieve(query_str) if len(relevant_nodes) > 0: table_row_context = "\nHere are some relevant example rows (values in the same order as columns above)\n" for node in relevant_nodes: table_row_context += str(node.get_content()) + "\n" table_info += table_row_context if verbose: print(f"> Table Info: {table_info}") context_strs.append(table_info) return "\n\n".join(context_strs)

定义扩展工作流¶

我们复用第1节中的工作流，但在文本到SQL生成后增加了升级版的SQL解析步骤。

通过子类化和扩展现有工作流来定制更高级的步骤非常简单。这里我们定义了一个新工作流，它重写了原有的retrieve_tables步骤以返回相关数据行。

In [ ]:

from llama_index.core.workflow import (
    Workflow,
    StartEvent,
    StopEvent,
    step,
    Context,
    Event,
)


class TextToSQLWorkflow2(TextToSQLWorkflow1):
    """Text-to-SQL Workflow that does query-time row AND table retrieval."""

    @step
    def retrieve_tables(
        self, ctx: Context, ev: StartEvent
    ) -> TableRetrieveEvent:
        """Retrieve tables."""
        table_schema_objs = self.obj_retriever.retrieve(ev.query)
        table_context_str = get_table_context_and_rows_str(
            ev.query, table_schema_objs, verbose=self._verbose
        )
        return TableRetrieveEvent(
            table_context_str=table_context_str, query=ev.query
        )

from llama_index.core.workflow import ( Workflow, StartEvent, StopEvent, step, Context, Event, ) class TextToSQLWorkflow2(TextToSQLWorkflow1): """Text-to-SQL Workflow that does query-time row AND table retrieval.""" @step def retrieve_tables( self, ctx: Context, ev: StartEvent ) -> TableRetrieveEvent: """Retrieve tables.""" table_schema_objs = self.obj_retriever.retrieve(ev.query) table_context_str = get_table_context_and_rows_str( ev.query, table_schema_objs, verbose=self._verbose ) return TableRetrieveEvent( table_context_str=table_context_str, query=ev.query )

由于整体步骤顺序相同，图表看起来应该也是一样的。

In [ ]:

from llama_index.utils.workflow import draw_all_possible_flows

draw_all_possible_flows(
    TextToSQLWorkflow2, filename="text_to_sql_table_retrieval.html"
)

from llama_index.utils.workflow import draw_all_possible_flows draw_all_possible_flows( TextToSQLWorkflow2, filename="text_to_sql_table_retrieval.html" )

text_to_sql_table_retrieval.html

执行查询操作¶

现在即使查询条件与数据库中的条目不完全匹配，我们也能检索到相关条目。

In [ ]:

workflow2 = TextToSQLWorkflow2(
    obj_retriever,
    text2sql_prompt,
    sql_retriever,
    response_synthesis_prompt,
    llm,
    verbose=True,
)

workflow2 = TextToSQLWorkflow2( obj_retriever, text2sql_prompt, sql_retriever, response_synthesis_prompt, llm, verbose=True, )

In [ ]:

response = await workflow2.run(
    query="What was the year that The Notorious BIG was signed to Bad Boy?"
)
print(str(response))

response = await workflow2.run( query="What was the year that The Notorious BIG was signed to Bad Boy?" ) print(str(response))

Running step retrieve_tables
VERBOSE: True
> Table Info: Table 'bad_boy_artists_album_release_summary' has columns: Act (VARCHAR), Year_signed (INTEGER), _Albums_released_under_Bad_Boy (VARCHAR), . The table description is: A summary of artists signed to Bad Boy Records along with the year they were signed and the number of albums they released.
Here are some relevant example rows (values in the same order as columns above)
('The Notorious B.I.G', 1993, '5')

> Table Info: Table 'filmography_of_diane_drummond' has columns: Year (INTEGER), Title (VARCHAR), Role (VARCHAR), Notes (VARCHAR), . The table description is: A list of film and television roles played by Diane Drummond from 1995 to 2001.
Here are some relevant example rows (values in the same order as columns above)
(2013, 'L.A. Slasher', 'The Actress', None)

> Table Info: Table 'progressive_rock_album_chart_positions' has columns: Year (INTEGER), Title (VARCHAR), Chart_Positions_UK (VARCHAR), Chart_Positions_US (VARCHAR), Chart_Positions_NL (VARCHAR), Comments (VARCHAR), . The table description is: Chart positions of progressive rock albums in the UK, US, and NL from 1969 to 1981.
Here are some relevant example rows (values in the same order as columns above)
(1977, 'Novella', '–', '46', '–', '1977 (January in US, August in UK, as the band moved to the Warner Bros Music Group)')

Step retrieve_tables produced event TableRetrieveEvent
Running step generate_sql
Step generate_sql produced event TextToSQLEvent
Running step generate_response
Step generate_response produced event StopEvent
assistant: The Notorious B.I.G. was signed to Bad Boy Records in 1993.