DataFocus

English

Overview

DataFocus includes two tools, FocusSQL and FocusGPT. FocusSQL is the Hallucination controllable Text2SQL component，FocusGPT is the fast response ChatBI.

There are already so many Text-to-SQL frameworks. Why do we still need another one?

In simple terms, FocusSQL adopts a two-step SQL generation solution, which enables control over the hallucinations of LLM and truly builds the trust of non-technical users in the generated SQL results.

Below is the comparison table between FocusSQL and others:

Comparison Analysis Table

Here’s a side-by-side comparison of DataFocus plugin with other LLM-based frameworks:

Feature	Traditional LLM Frameworks	FocusSQL
Generation Process	Black box, direct SQL generation	Transparent, two-step (keywords + SQL)
Hallucination Risk	High, depends on model quality	Low, controllable (keyword verification)
Speed	Slow, relies on large model inference	Fast, deterministic keyword-to-SQL
Cost	High, requires advanced models	Low, reduces reliance on large models
Non-Technical User Friendliness	Low, hard to verify results	High, easy keyword checking

The following will introduce how to configure and an example demonstration.

1. Apply for DataFocus Token

If you don't have the DataFocus application yet, please apply for one on the DataFocus Website.
Log in to your DataFocus application. Click Admin > Interface Authentication > Bearer Token > New Bearer Token, to create a new token and get the token value.

If you have a DataFocus private deployment environment, you can get Token on your own environment.

2. Fill in the configuration in Dify

Install DataFocus from Marketplace and fill token and host in the authorization page.Token is the value obtained in the previous step.If you have a DataFocus private deployment environment, host is your environment host. Otherwise, the SAAS environment address can be used by default.

3. Use the tool

DataFocus includes two tools, FocusSQL and FocusGPT.

FocusSQL

FocusSQL is a natural language to SQL plugin based on keyword parsing.

Output Variable JSON

Name	Type	Description
content	string	Generated SQL statements
question	string	Generated keywords
type	string	Return type

Output Example

FocusGPT

FocusGPT is an intelligent query plugin that supports multiple rounds of conversations, which allow you query data from your database.
FocusGPT not only can return query SQL but also return query result to you.

Output Variable JSON

name	type	Description
code	number	Status code
columns	[[object]]	Two-dimensional array storing query results
count	number	Number of rows returned
duration	string	Query execution time, in seconds (s)
headers	[object]	Column header information for the two-dimensional array columns
» display	string	Display name of the column header
» name	string	Original column name of the header
» suf	string	Prefix of the column header, indicating aggregation method
sql	[object]	SQL corresponding to the query data
»select_clause	string	SQL corresponding to the query data
title	string	Keywords generated from parsing

Output Example

Configuration

FocusSQL and FocusGPT have similar configuration. Below are the functions and usage instructions of each parameter

Parameter	Description
Language	Language environment, only support Chinese and English
Query Statement	Natural language input by users
Table Name	Target data table for query
Data Model	Custom model parameter entry
Output SQL Type	Output SQL Type
Conversation Id	Unique identifier of the session, which allow tool identify and maintain session state
Action	The behavior of tool execution currently includes two types: obtaining table lists and dialogues
Datasource Type	Types of external data sources connected. If datasource type was selected, the connection parameters below need to be filled in
Host	host
Port	port
DB user	user
DB Password	password
Database Name	database name
JDBC	JDBC
Schema	Schema name

Model Parameters

The data model needs to pass in a JSON string, and the structure of the model is as follows

Structure

Name	Type	Required	Description
type	string	Yes	Database type
version	string	Yes	Database version, eg: 8.0
tables	[object]	Yes	Table structure list
» tableDisplayName	string	No	Table display name
» tableName	string	No	Original table name
» columns	[object]	No	Columns structure list
»» columnDisplayName	string	Yes	Column display name
»» columnName	string	Yes	Original column name
»» dataType	string	Yes	Column data type
»» aggregation	string	Yes	Column default aggregation
relations	[object]	Yes	Association relationship list
» conditions	[object]	No	Associated conditions
»» dstColName	string	No	Dimension original column name
»» srcColName	string	No	Fact original column name
» dimensionTable	string	No	Dimension original table name
» factTable	string	No	Fact original table name
» joinType	string	No	Association type

Parameter values

type

DataBase	Value
MySQL	mysql
ClickHouse	clickhouse
Impala	impala

dataType

DataType	Value
Boolean	boolean
Integer	int
Long integer	bigint
Float	double
String	string
Timestamp	timestamp
Date type	date
Time type	time

aggregation

Aggregation	Value
Sum	SUM
Mean	AVERAGE
Min	MIN
Max	MAX
Count	COUNT
Number of deduplicates	COUNT_DISTINCT
Variance	VARIANCE
Standard deviation	STD_DEVIATION
None	NONE

joinType

Assocation	Value
Left association	LEFT JOIN
Right association	RIGHT JOIN
Internal	INNER JOIN
Fully associative	FULL JOIN

Example

model

Consult

WeChat: datafocus2018

DataFocus Discord