Under the hood, LangChain uses SQLAlchemy to connect to SQL databases. The `SQLDatabaseChain` can therefore be used with any SQL dialect supported by SQLAlchemy, such as MS SQL, MySQL, MariaDB, PostgreSQL, Oracle SQL, [Databricks](/docs/ecosystem/integrations/databricks.html) and SQLite. Please refer to the SQLAlchemy documentation for more information about requirements for connecting to your database. For example, a connection to MySQL requires an appropriate connector such as PyMySQL. A URI for a MySQL connection might look like: `mysql+pymysql://user:pass@some_mysql_db_address/db_name`.
This demonstration uses SQLite and the example Chinook database.
To set it up, follow the instructions on https://database.guide/2-sample-databases-sqlite/, placing the `.db` file in a notebooks folder at the root of this repository.
db = SQLDatabase.from_uri("sqlite:///../../../../notebooks/Chinook.db")
llm = OpenAI(temperature=0, verbose=True)
```
**NOTE:** For data-sensitive projects, you can specify `return_direct=True` in the `SQLDatabaseChain` initialization to directly return the output of the SQL query without any additional formatting. This prevents the LLM from seeing any contents within the database. Note, however, the LLM still has access to the database scheme (i.e. dialect, table and key names) by default.
/workspace/langchain/langchain/sql_database.py:191: SAWarning: Dialect sqlite+pysqlite does *not* support Decimal objects natively, and SQLAlchemy must convert from floating point - rounding errors and other issues may occur. Please consider storing Decimal numbers as strings or integers on this platform for lossless storage.
sample_rows = connection.execute(command)
SELECT COUNT(*) FROM "Employee";
SQLResult: [(8,)]
Answer:There are 8 employees.
> Finished chain.
'There are 8 employees.'
```
</CodeOutputBlock>
## Use Query Checker
Sometimes the Language Model generates invalid SQL with small mistakes that can be self-corrected using the same technique used by the SQL Database Agent to try and fix the SQL using the LLM. You can simply specify this option when creating the chain:
SQLQuery:SELECT COUNT(*) FROM Album WHERE ArtistId = 3;
SQLResult: [(1,)]
Answer:There is 1 album by Aerosmith.
> Finished chain.
'There is 1 album by Aerosmith.'
```
</CodeOutputBlock>
## Customize Prompt
You can also customize the prompt that is used. Here is an example prompting it to understand that foobar is the same as the Employee table
```python
from langchain.prompts.prompt import PromptTemplate
_DEFAULT_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.
Use the following format:
Question: "Question here"
SQLQuery: "SQL Query to run"
SQLResult: "Result of the SQLQuery"
Answer: "Final answer here"
Only use the following tables:
{table_info}
If someone asks for the table foobar, they really mean the employee table.
db_chain.run("How many employees are there in the foobar table?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseChain chain...
How many employees are there in the foobar table?
SQLQuery:SELECT COUNT(*) FROM Employee;
SQLResult: [(8,)]
Answer:There are 8 employees in the foobar table.
> Finished chain.
'There are 8 employees in the foobar table.'
```
</CodeOutputBlock>
## Return Intermediate Steps
You can also return the intermediate steps of the SQLDatabaseChain. This allows you to access the SQL statement that was generated, as well as the result of running that against the SQL Database.
from langchain_experimental.sql import SQLDatabaseChain
```
Set up the SQLDatabase and LLM
```python
db = SQLDatabase.from_uri("sqlite:///../../../../notebooks/Chinook.db")
llm = OpenAI(temperature=0, verbose=True)
```
Set up the memory
```python
from langchain.memory import ConversationBufferMemory
memory = ConversationBufferMemory()
```
Now we need to add a place for memory in the prompt template
```python
from langchain.prompts import PromptTemplate
PROMPT_SUFFIX = """Only use the following tables:
{table_info}
Previous Conversation:
{history}
Question: {input}"""
_DEFAULT_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. Unless the user specifies in his question a specific number of examples he wishes to obtain, always limit your query to at most {top_k} results. 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 the 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. Also, pay attention to which column is in which table.
## Choosing how to limit the number of rows returned
If you are querying for several rows of a table you can select the maximum number of results you want to get by using the 'top_k' parameter (default is 10). This is useful for avoiding query results that exceed the prompt max length or consume tokens unnecessarily.
db_chain.run("What are some example tracks by composer Johann Sebastian Bach?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseChain chain...
What are some example tracks by composer Johann Sebastian Bach?
SQLQuery:SELECT Name FROM Track WHERE Composer = 'Johann Sebastian Bach' LIMIT 3
SQLResult: [('Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace',), ('Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria',), ('Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude',)]
Answer:Examples of tracks by Johann Sebastian Bach are Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace, Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria, and Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude.
> Finished chain.
'Examples of tracks by Johann Sebastian Bach are Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace, Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria, and Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude.'
```
</CodeOutputBlock>
## Adding example rows from each table
Sometimes, the format of the data is not obvious and it is optimal to include a sample of rows from the tables in the prompt to allow the LLM to understand the data before providing a final query. Here we will use this feature to let the LLM know that artists are saved with their full names by providing two rows from the `Track` table.
```python
db = SQLDatabase.from_uri(
"sqlite:///../../../../notebooks/Chinook.db",
include_tables=['Track'], # we include only one table to save tokens in the prompt :)
sample_rows_in_table_info=2)
```
The sample rows are added to the prompt after each corresponding table's column information:
db_chain.run("What are some example tracks by Bach?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseChain chain...
What are some example tracks by Bach?
SQLQuery:SELECT "Name", "Composer" FROM "Track" WHERE "Composer" LIKE '%Bach%' LIMIT 5
SQLResult: [('American Woman', 'B. Cummings/G. Peterson/M.J. Kale/R. Bachman'), ('Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace', 'Johann Sebastian Bach'), ('Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria', 'Johann Sebastian Bach'), ('Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude', 'Johann Sebastian Bach'), ('Toccata and Fugue in D Minor, BWV 565: I. Toccata', 'Johann Sebastian Bach')]
Answer:Tracks by Bach include 'American Woman', 'Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace', 'Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria', 'Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude', and 'Toccata and Fugue in D Minor, BWV 565: I. Toccata'.
> Finished chain.
'Tracks by Bach include \'American Woman\', \'Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace\', \'Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria\', \'Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude\', and \'Toccata and Fugue in D Minor, BWV 565: I. Toccata\'.'
```
</CodeOutputBlock>
### Custom Table Info
In some cases, it can be useful to provide custom table information instead of using the automatically generated table definitions and the first `sample_rows_in_table_info` sample rows. For example, if you know that the first few rows of a table are uninformative, it could help to manually provide example rows that are more diverse or provide more information to the model. It is also possible to limit the columns that will be visible to the model if there are unnecessary columns.
This information can be provided as a dictionary with table names as the keys and table information as the values. For example, let's provide a custom definition and sample rows for the Track table with only a few columns:
```python
custom_table_info = {
"Track": """CREATE TABLE Track (
"TrackId" INTEGER NOT NULL,
"Name" NVARCHAR(200) NOT NULL,
"Composer" NVARCHAR(220),
PRIMARY KEY ("TrackId")
)
/*
3 rows from Track table:
TrackId Name Composer
1 For Those About To Rock (We Salute You) Angus Young, Malcolm Young, Brian Johnson
2 Balls to the Wall None
3 My favorite song ever The coolest composer of all time
*/"""
}
```
```python
db = SQLDatabase.from_uri(
"sqlite:///../../../../notebooks/Chinook.db",
include_tables=['Track', 'Playlist'],
sample_rows_in_table_info=2,
custom_table_info=custom_table_info)
print(db.table_info)
```
<CodeOutputBlock lang="python">
```
CREATE TABLE "Playlist" (
"PlaylistId" INTEGER NOT NULL,
"Name" NVARCHAR(120),
PRIMARY KEY ("PlaylistId")
)
/*
2 rows from Playlist table:
PlaylistId Name
1 Music
2 Movies
*/
CREATE TABLE Track (
"TrackId" INTEGER NOT NULL,
"Name" NVARCHAR(200) NOT NULL,
"Composer" NVARCHAR(220),
PRIMARY KEY ("TrackId")
)
/*
3 rows from Track table:
TrackId Name Composer
1 For Those About To Rock (We Salute You) Angus Young, Malcolm Young, Brian Johnson
2 Balls to the Wall None
3 My favorite song ever The coolest composer of all time
*/
```
</CodeOutputBlock>
Note how our custom table definition and sample rows for `Track` overrides the `sample_rows_in_table_info` parameter. Tables that are not overridden by `custom_table_info`, in this example `Playlist`, will have their table info gathered automatically as usual.
db_chain.run("What are some example tracks by Bach?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseChain chain...
What are some example tracks by Bach?
SQLQuery:SELECT "Name" FROM Track WHERE "Composer" LIKE '%Bach%' LIMIT 5;
SQLResult: [('American Woman',), ('Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace',), ('Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria',), ('Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude',), ('Toccata and Fugue in D Minor, BWV 565: I. Toccata',)]
Answer:text='You are a SQLite expert. Given an input question, first create a syntactically correct SQLite query to run, then look at the results of the query and return the answer to the input question.\nUnless the user specifies in the question a specific number of examples to obtain, query for at most 5 results using the LIMIT clause as per SQLite. You can order the results to return the most informative data in the database.\nNever query for all columns from a table. You must query only the columns that are needed to answer the question. Wrap each column name in double quotes (") to denote them as delimited identifiers.\nPay attention to use only the column names you can see in the tables below. Be careful to not query for columns that do not exist. Also, pay attention to which column is in which table.\n\nUse the following format:\n\nQuestion: "Question here"\nSQLQuery: "SQL Query to run"\nSQLResult: "Result of the SQLQuery"\nAnswer: "Final answer here"\n\nOnly use the following tables:\n\nCREATE TABLE "Playlist" (\n\t"PlaylistId" INTEGER NOT NULL, \n\t"Name" NVARCHAR(120), \n\tPRIMARY KEY ("PlaylistId")\n)\n\n/*\n2 rows from Playlist table:\nPlaylistId\tName\n1\tMusic\n2\tMovies\n*/\n\nCREATE TABLE Track (\n\t"TrackId" INTEGER NOT NULL, \n\t"Name" NVARCHAR(200) NOT NULL,\n\t"Composer" NVARCHAR(220),\n\tPRIMARY KEY ("TrackId")\n)\n/*\n3 rows from Track table:\nTrackId\tName\tComposer\n1\tFor Those About To Rock (We Salute You)\tAngus Young, Malcolm Young, Brian Johnson\n2\tBalls to the Wall\tNone\n3\tMy favorite song ever\tThe coolest composer of all time\n*/\n\nQuestion: What are some example tracks by Bach?\nSQLQuery:SELECT "Name" FROM Track WHERE "Composer" LIKE \'%Bach%\' LIMIT 5;\nSQLResult: [(\'American Woman\',), (\'Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace\',), (\'Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria\',), (\'Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude\',), (\'Toccata and Fugue in D Minor, BWV 565: I. Toccata\',)]\nAnswer:'
You are a SQLite expert. Given an input question, first create a syntactically correct SQLite query to run, then look at the results of the query and return the answer to the input question.
Unless the user specifies in the question a specific number of examples to obtain, query for at most 5 results using the LIMIT clause as per SQLite. You can order the results to return the most informative data in the database.
Never query for all columns from a table. You must query only the columns that are needed to answer the question. Wrap each column name in double quotes (") to denote them as delimited identifiers.
Pay attention to use only the column names you can see in the tables below. Be careful to not query for columns that do not exist. Also, pay attention to which column is in which table.
Use the following format:
Question: "Question here"
SQLQuery: "SQL Query to run"
SQLResult: "Result of the SQLQuery"
Answer: "Final answer here"
Only use the following tables:
CREATE TABLE "Playlist" (
"PlaylistId" INTEGER NOT NULL,
"Name" NVARCHAR(120),
PRIMARY KEY ("PlaylistId")
)
/*
2 rows from Playlist table:
PlaylistId Name
1 Music
2 Movies
*/
CREATE TABLE Track (
"TrackId" INTEGER NOT NULL,
"Name" NVARCHAR(200) NOT NULL,
"Composer" NVARCHAR(220),
PRIMARY KEY ("TrackId")
)
/*
3 rows from Track table:
TrackId Name Composer
1 For Those About To Rock (We Salute You) Angus Young, Malcolm Young, Brian Johnson
2 Balls to the Wall None
3 My favorite song ever The coolest composer of all time
*/
Question: What are some example tracks by Bach?
SQLQuery:SELECT "Name" FROM Track WHERE "Composer" LIKE '%Bach%' LIMIT 5;
SQLResult: [('American Woman',), ('Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace',), ('Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria',), ('Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude',), ('Toccata and Fugue in D Minor, BWV 565: I. Toccata',)]
Answer:
{'input': 'What are some example tracks by Bach?\nSQLQuery:SELECT "Name" FROM Track WHERE "Composer" LIKE \'%Bach%\' LIMIT 5;\nSQLResult: [(\'American Woman\',), (\'Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace\',), (\'Aria Mit 30 Veränderungen, BWV 988 "Goldberg Variations": Aria\',), (\'Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude\',), (\'Toccata and Fugue in D Minor, BWV 565: I. Toccata\',)]\nAnswer:', 'top_k': '5', 'dialect': 'sqlite', 'table_info': '\nCREATE TABLE "Playlist" (\n\t"PlaylistId" INTEGER NOT NULL, \n\t"Name" NVARCHAR(120), \n\tPRIMARY KEY ("PlaylistId")\n)\n\n/*\n2 rows from Playlist table:\nPlaylistId\tName\n1\tMusic\n2\tMovies\n*/\n\nCREATE TABLE Track (\n\t"TrackId" INTEGER NOT NULL, \n\t"Name" NVARCHAR(200) NOT NULL,\n\t"Composer" NVARCHAR(220),\n\tPRIMARY KEY ("TrackId")\n)\n/*\n3 rows from Track table:\nTrackId\tName\tComposer\n1\tFor Those About To Rock (We Salute You)\tAngus Young, Malcolm Young, Brian Johnson\n2\tBalls to the Wall\tNone\n3\tMy favorite song ever\tThe coolest composer of all time\n*/', 'stop': ['\nSQLResult:']}
Examples of tracks by Bach include "American Woman", "Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace", "Aria Mit 30 Veränderungen, BWV 988 'Goldberg Variations': Aria", "Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude", and "Toccata and Fugue in D Minor, BWV 565: I. Toccata".
> Finished chain.
'Examples of tracks by Bach include "American Woman", "Concerto for 2 Violins in D Minor, BWV 1043: I. Vivace", "Aria Mit 30 Veränderungen, BWV 988 \'Goldberg Variations\': Aria", "Suite for Solo Cello No. 1 in G Major, BWV 1007: I. Prélude", and "Toccata and Fugue in D Minor, BWV 565: I. Toccata".'
In some case, the table schema can be hidden behind a JSON or JSONB column. Adding row samples into the prompt might help won't always describe the data perfectly.
chain.run("How many employees are also customers?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseSequentialChain chain...
Table names to use:
['Employee', 'Customer']
> Entering new SQLDatabaseChain chain...
How many employees are also customers?
SQLQuery:SELECT COUNT(*) FROM Employee e INNER JOIN Customer c ON e.EmployeeId = c.SupportRepId;
SQLResult: [(59,)]
Answer:59 employees are also customers.
> Finished chain.
> Finished chain.
'59 employees are also customers.'
```
</CodeOutputBlock>
## Using Local Language Models
Sometimes you may not have the luxury of using OpenAI or other service-hosted large language model. You can, ofcourse, try to use the `SQLDatabaseChain` with a local model, but will quickly realize that most models you can run locally even with a large GPU struggle to generate the right output.
```python
import logging
import torch
from transformers import AutoTokenizer, GPT2TokenizerFast, pipeline, AutoModelForSeq2SeqLM, AutoModelForCausalLM
/workspace/langchain/.venv/lib/python3.9/site-packages/tqdm/auto.py:21: TqdmWarning: IProgress not found. Please update jupyter and ipywidgets. See https://ipywidgets.readthedocs.io/en/stable/user_install.html
This model should work for very simple SQL queries, as long as you use the query checker as specified above, e.g.:
```python
local_chain("How many customers are there?")
```
<CodeOutputBlock lang="python">
```
> Entering new SQLDatabaseChain chain...
How many customers are there?
SQLQuery:
/workspace/langchain/.venv/lib/python3.9/site-packages/transformers/pipelines/base.py:1070: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
warnings.warn(
/workspace/langchain/.venv/lib/python3.9/site-packages/transformers/pipelines/base.py:1070: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
warnings.warn(
SELECT count(*) FROM Customer
SQLResult: [(59,)]
Answer:
/workspace/langchain/.venv/lib/python3.9/site-packages/transformers/pipelines/base.py:1070: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
warnings.warn(
[59]
> Finished chain.
{'query': 'How many customers are there?',
'result': '[59]',
'intermediate_steps': [{'input': 'How many customers are there?\nSQLQuery:SELECT count(*) FROM Customer\nSQLResult: [(59,)]\nAnswer:',
'top_k': '5',
'dialect': 'sqlite',
'table_info': '\nCREATE TABLE "Customer" (\n\t"CustomerId" INTEGER NOT NULL, \n\t"FirstName" NVARCHAR(40) NOT NULL, \n\t"LastName" NVARCHAR(20) NOT NULL, \n\t"Company" NVARCHAR(80), \n\t"Address" NVARCHAR(70), \n\t"City" NVARCHAR(40), \n\t"State" NVARCHAR(40), \n\t"Country" NVARCHAR(40), \n\t"PostalCode" NVARCHAR(10), \n\t"Phone" NVARCHAR(24), \n\t"Fax" NVARCHAR(24), \n\t"Email" NVARCHAR(60) NOT NULL, \n\t"SupportRepId" INTEGER, \n\tPRIMARY KEY ("CustomerId"), \n\tFOREIGN KEY("SupportRepId") REFERENCES "Employee" ("EmployeeId")\n)\n\n/*\n3 rows from Customer table:\nCustomerId\tFirstName\tLastName\tCompany\tAddress\tCity\tState\tCountry\tPostalCode\tPhone\tFax\tEmail\tSupportRepId\n1\tLuís\tGonçalves\tEmbraer - Empresa Brasileira de Aeronáutica S.A.\tAv. Brigadeiro Faria Lima, 2170\tSão José dos Campos\tSP\tBrazil\t12227-000\t+55 (12) 3923-5555\t+55 (12) 3923-5566\tluisg@embraer.com.br\t3\n2\tLeonie\tKöhler\tNone\tTheodor-Heuss-Straße 34\tStuttgart\tNone\tGermany\t70174\t+49 0711 2842222\tNone\tleonekohler@surfeu.de\t5\n3\tFrançois\tTremblay\tNone\t1498 rue Bélanger\tMontréal\tQC\tCanada\tH2G 1A7\t+1 (514) 721-4711\tNone\tftremblay@gmail.com\t3\n*/',
'stop': ['\nSQLResult:']},
'SELECT count(*) FROM Customer',
{'query': 'SELECT count(*) FROM Customer', 'dialect': 'sqlite'},
Even this relatively large model will most likely fail to generate more complicated SQL by itself. However, you can log its inputs and outputs so that you can hand-correct them and use the corrected examples for few-shot prompt examples later. In practice, you could log any executions of your chain that raise exceptions (as shown in the example below) or get direct user feedback in cases where the results are incorrect (but did not raise an exception).
List all the customer first names that start with 'a'
SQLQuery:
/workspace/langchain/.venv/lib/python3.9/site-packages/transformers/pipelines/base.py:1070: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
warnings.warn(
SELECT firstname FROM customer WHERE firstname LIKE '%a%'
/workspace/langchain/.venv/lib/python3.9/site-packages/transformers/pipelines/base.py:1070: UserWarning: You seem to be using the pipelines sequentially on GPU. In order to maximize efficiency please use a dataset
Tremblay\tNone\t1498 rue Bélanger\tMontréal\tQC\tCanada\tH2G 1A7\t+1 (514) 721-4711\t\
None\tftremblay@gmail.com\t3\n*/"
```
</CodeOutputBlock>
Run the snippet above a few times, or log exceptions in your deployed environment, to collect lots of examples of inputs, table_info and sql_cmd generated by your language model. The sql_cmd values will be incorrect and you can manually fix them up to build a collection of examples, e.g. here we are using YAML to keep a neat record of our inputs and corrected SQL output that we can build up over time.
