SaaS Semantic SQL Analytics Platform

Yitu AI Data Analytics

AI-Powered Semantic SQL Analytics Platform

Yitu AI Data Analytics is a SaaS-based semantic SQL analytics platform that enables business users to query and analyze databases using natural language. Through the Semantic Modeling Layer (MDL), physical table structures are mapped to business terms, allowing non-technical users to perform complex data analysis without SQL skills.

SQL Analysis

> What were the top 10 products by sales in East China last month?

SELECT product_name, SUM(amount) as total

FROM orders WHERE region = 'East China'

AND date >= '2026-04-01'

GROUP BY product_name

ORDER BY total DESC LIMIT 10;

22+
Data Sources
95%+
SQL Accuracy
<3 sec
Query Response
50M+
Annual Queries

How It Works

Three steps to complete data analysis

1

Connect Data Source

Connect your database or upload CSV files. The system auto-discovers table structures and generates a semantic model

2

Ask in Plain Language

Enter a business question like 'What were the top 10 products by sales in East China last month?'

3

Get Analysis Results

AI generates SQL, runs security checks, executes the query, and returns visualized results with insights

Features

From data source connection to intelligent analysis, Yitu provides a complete data analytics workflow

AI Natural Language Query

Describe your analysis in plain language, AI understands semantics and generates precise SQL — zero barrier to data querying

22+ Data Sources

Supports PostgreSQL, MySQL, BigQuery, Snowflake and other major databases and cloud data warehouses

Semantic Modeling (MDL)

Maps physical table structures to business terms, enabling AI to accurately understand data meaning and improve SQL accuracy

Enterprise-Grade Security

RBAC multi-tenant permissions, SQL security policy review, row/column-level access control, multi-layer database credential encryption

Sub-Second Response

Real-time AI conversational analysis, SSE streaming results — from question to visualization in seconds

Semantic Memory System

Vector database-based query memory, similar questions reuse successful mappings — gets smarter with use

Multi-Dimensional Analysis

Extracts Cube definitions from MDL, supports metric/dimension combo queries, auto-generates aggregate SQL

SQL Editor

Professional editor for technical users with syntax highlighting, execution plan preview, and async queries

Enterprise Security

RBAC permission system, SQL safety policies, row/column-level access control, multi-layer database credential encryption

Broad Compatibility

22+ data source support, from traditional databases to cloud data warehouses — all in one place

Lightning Fast

SSE streaming results, real-time AI conversational analysis — from question to chart in seconds

FAQ

Why business users can now self-serve reports

Yitu AI Data Analytics turns 'ask one question, get one chart' into the default experience for business users. Data teams stop being a ticket funnel; business users stop needing SQL to answer a number. A Semantic Modeling Layer (MDL) maps physical tables to business terms, an LLM produces safety-checked SQL against that layer, and results stream back via SSE with charts — auditable, reversible, reusable. Customers in retail, cross-border e-commerce, SaaS, and financial risk have compressed data-request cycles from days to seconds and cut repeat ticket volume by 70%+ within three months.

Three real scenarios

Retail chain · Regional ops

Situation · Needs top 10 SKUs by revenue in East China last month, YoY.

Outcome · Asks in the Yitu chat; SQL is generated with region/date filters, executed, and charted in under 30 seconds. A follow-up 'now re-rank by margin' takes one more sentence.

Cross-border e-commerce · Growth

Situation · Weekly ROAS/conversion/AOV report to the ads team, repetitive and prone to definition drift.

Outcome · Common questions are saved as MDL Cubes; semantic memory reuses successful mappings; the analyst clicks 'Channel Weekly' and gets a consistent result set to export or push to Feishu.

Financial risk · Compliance audit

Situation · Needs T+1 anomaly checks across databases with row/column-level permissions; DBAs won't issue ad-hoc accounts.

Outcome · Strict SQL mode blocks writes; row/column permissions bind to RBAC; auditors only see authorized data; every query and generated SQL is logged and replayable.

Performance and scale

22+Data sources

PG/MySQL/Oracle/BigQuery/Snowflake/Redshift/ClickHouse/Databricks and more

95%+SQL accuracy

Blind test on 5,000 business questions; failed queries auto-retry

<3 secMedian query response

From prompt to first SSE chunk, includes LLM inference

50M+Annual queries

Aggregate SQL executions across 2025 for customers in production

70%↓Ticket volume drop

Data-team tickets after 3 months vs historical average

≤1 dayOn-prem deployment

Helm Chart with HA replicas

How we compare

vs traditional BI (Tableau / PowerBI / FanRuan)
Us

Natural-language queries, MDL keeps definitions consistent.

Them

Drag-and-drop fields, DAX, scattered logic; business users still depend on data teams.

Takeaway · BI is deep reporting; Yitu is daily self-serve. They coexist.

vs generic ChatBI / Text2SQL projects
Us

MDL + semantic memory + SQL safety policy; enterprise-grade permissions built in.

Them

Usually 'prompt-stuffed schemas' — accuracy collapses on multi-table JOINs and permission constraints.

Takeaway · The gap from demo to production is the semantic layer and safety policy.

vs building on LLM + LangChain
Us

SaaS or on-prem, 2 weeks to value, ongoing upgrades by us.

Them

Selection, vector store, safety, permission middleware, alerting — 6+ engineer-months, then perpetual maintenance.

Takeaway · First-time working is easy; running it for three years without incidents is the actual cost.

Deeper details

How the MDL semantic layer works

MDL maps physical tables like orders/order_items/customers to business terms such as Order, Order Line, Customer, and explicitly declares primary keys, foreign keys, dimensions, measures, and definitions. The LLM reads MDL before generating SQL, eliminating ghost errors like 'is amount pre-tax or post-tax'. One company, one definition — that is what democratization actually requires.

SQL safety and permission control

All writes are rejected by default; SELECTs pass through RBAC checks with row-level filters and column-level masking injected before dispatch. Credentials are Fernet-encrypted and decrypted only at execution time. Every query and generated SQL is logged and replayable per user/role/time — meeting internal audits, external audits, GB/T 22239 level 3, and financial-industry review.

Semantic memory: better with use

Every query marked 'useful' by a business user is stored — its natural-language prompt, generated SQL, tables touched — in a vector store. Similar future questions retrieve candidates first and let the LLM rewrite, pushing accuracy from a cold-start ~80% to sustained 95%+ and cutting token cost by ~40%. Customers buy not a one-time SQL-generation feature but an analyst that keeps learning their business.

Rollout path

Day 1: SaaS trial with your database; Free plan's 100 monthly queries prove 3-5 high-frequency questions. Week 1: build MDL, load 5-10 core Cubes, business users go self-serve. Month 1: SSO/RBAC, training, formal rollout. Month 3: most customers upgrade to Pro/Team and retire 'data request' tickets. On-prem customers get Helm Chart plus a dedicated CSM, HA cluster in two weeks.

Redefine Data Analytics with AI

No SQL needed — ask questions in plain language and get complex analysis done