Quick Start with Alibaba Cloud ADA: Your First Conversational Data Analysis

About AI-Native Big Data Service ADA

ADA integrates Agent technology with big data and AI engines (MaxCompute, Hologres, Flink, EMR, DataWorks) to provide a next-generation data productivity platform. Its core architecture features three capabilities:

Multi-agent collaborative orchestration : Ada acts as the master agent, handling intent understanding, task routing, and automatic coordination of specialized engine agents for cross-engine data delivery.

Enterprise-grade knowledge engine : Built-in semantic graph capabilities (same source as DataWorks) unify data assets, business semantics, and analytical experience into a cognitive hub.

High-flexibility custom agents : Supports both low-code and high-code modes, allowing enterprises to plug in mainstream agent engines and build domain-specific agents.

Business Scenario & Pain Points

Operations teams often ask: "What are the last 7 days' order volumes per region, their week-over-week changes, which region fluctuates most, and why?" Traditional workflow: ops submits request → data team schedules, writes SQL, exports data → insights arrive too late for decisions. Adding dimensions requires new requests, incurring high communication and time costs.

Solution Architecture

Two parallel entry points:

Portal Console : Service activation, version subscription, usage management.

AdaWorks Workbench : Data ingestion, conversational execution, result inspection.

Data ingestion method can be switched in the workbench without affecting subsequent conversational steps.

Implementation Steps (≈10 minutes end-to-end)

Step 1: Confirm Service Availability

Goal: Verify account permissions and select operating region.

Log in to Alibaba Cloud console, open ADA console, choose target region (must match downstream engine instances). New users can claim free trial quota. Console URL: https://console.aliyun.com/ada/cn-hangzhou/console/overview On the Overview page, subscribe via the AdaWork Version Subscription card (trial users may skip).

Under Agent Capabilities, select Ada to start chatting.

✅ Done when : Chat UI loads and service status shows available.

Step 2: Bring Data into the Conversation

Goal: Give the agent analyzable data.

Open AdaWorks Workbench, click New Task in sidebar.

Click Add Content → Local File , upload the order detail spreadsheet.

Data requirements : Columns must include order time, region, city, channel, amount. Date column in standard format (e.g., 2026-09-06) covering at least two full periods for week-over-week calculation. Sample data : Download "示例数据-区域订单明细.xlsx" (14 days, 4 regions, 5 channels, standardized) via the "阅读原文" link.

✅ Done when : File uploads successfully and appears as added content in the input box.

Step 3: Initiate Analysis & Verify Process

Goal: Locate the problem and validate the execution path and computation, not just get a number.

Input question in natural language, no pre-defined dimensions needed. Example:

Compare order volume of last 7 days vs previous week, list week-over-week changes by region

.

Verify execution plan : Ada generates a plan; check data scope and breakdown dimensions. Confirm and click Continue. If plan mismatches expectations, add constraints directly in chat — no session restart needed.

Verify execution process : During run, collapsed sections (Processing/Processed) appear above the reply. Expand to see actual code and intermediate results. Data access triggers a permission card (Allow once / Allow always / Deny); first run choose "Allow once".

Verify computation logic : After reply, cross-check key figures in the conclusion against the execution steps to ensure statistical caliber and time range match expectations.

✅ Done when : Session produces a reply with data tables and analysis conclusions, and tool-call cards are expandable for audit.

Reference conclusion (using sample data) : Total volume down ~5% week-over-week. Drilling into regions reveals East China down >10% (only region with significant drop), others flat. This illustrates "totals often mask structural issues."

Step 4: Drill Down to Root Cause

Goal: Decompose macro issue into actionable micro granularity.

Previous step pinned "East China decline". Continue in same session (Ada inherits context automatically):

Break down reasons for East China order decline, list week-over-week changes by city and product category

Reference conclusion : Decline concentrated in Hangzhou (Shanghai, Suzhou flat), almost entirely from "Home Appliance" category. Report may note "Mini-program channel largest drop" but that's a symptom (Hangzhou appliances mostly ordered via mini-program); core issue is "Home Appliance decline".

Further confirm whether demand shrank or supply failed:

For Hangzhou home appliances, is it fewer orders placed or more cancellations?

Reference conclusion : Both fewer placed orders and higher cancellations appear simultaneously, cancellation rate significantly above other categories. Combined signal points to "stock-out for that category" rather than pure demand drop.

✅ Done when : Reply pinpoints decline to specific city and product, with preliminary attribution.

Tip : Tables and report cards support online preview and download. For long-term retention or cross-team sharing, convert to "Artifacts".

Step 5: Cement as Daily Automated Report

Goal: Turn one-off manual analysis into a daily scheduled job.

In Workbench sidebar go to Scheduled Tasks → New Scheduled Task , configure the validated analysis logic as a recurring task.

Configure message delivery : After auto-run, push conclusions to the team. Set a group bot webhook (DingTalk, Lark, WeCom — determined by URL) or pick IM bot with specific recipients/groups. Team receives anomaly alerts in daily chat tools.

✅ Done when : Task appears in "My Scheduled Tasks" with status "Running" and next run time. If not triggered on time, check "Execution Records" for status and errors.

Alternative Data Ingestion Methods (Replace Step 2)

If data already lives in a data warehouse or metric definitions are documented, use these advanced approaches; Steps 3+ remain identical.

Scenario A: Data in MaxCompute, Hologres, etc.

Use Connectors for direct query. A connector is the ingestion config between ADA and an engine instance; register once, reuse long-term.

Path : Sidebar → Extensions → Connectors → Register Connector. Service type "Big Data Engine", confirm region, pick instance, fill data location and credentials.

Key tips :

Description must be precise : e.g., "E-commerce ops order detail for regional order analysis". Agent uses this to decide when to call; vague descriptions cause mis-selection.

Region consistency : If instance list empty, verify console's current region.

After saving, mount the connector via "Add Content" in New Task.

Scenario B: Metric Definitions Scattered in Business Docs

Build a Knowledge Base to unify calibers, useful when same metric (e.g., "active user") has different definitions across departments.

Path : Sidebar → Knowledge Base → New Knowledge Base, upload metric definitions, region segmentation rules, etc. Wait for Wiki build, then bind in "Add Content".

Key tips :

After binding, Ada's replies carry Wiki citations (expandable to see exact caliber definitions) — absent when only uploading tables.

Knowledge base supplies "caliber", not "business data"; must pair with "upload table" or "connector", and prompt must explicitly say "calculate per knowledge base caliber".

Summary & Next Steps

You have completed four loops: data ingestion, conversational analysis, layered drill-down, automation hardening — closing the loop from "business question" to "auditable conclusion".

Note : Direct table upload suits quick validation, but metric accuracy depends on knowledge asset completeness. Complex metrics inferred solely from table structure may deviate, and conclusions won't carry caliber citations.

Recommended next directions per business priority:

Standardize metric calibers : Load high-frequency metric definitions into Knowledge Base so conclusions auto-cite calibers, boosting accuracy.

Direct-connect enterprise data warehouse : Replace local spreadsheet with engine instance direct connection for real-time analysis at scale.

Extend agent capabilities : Subscribe official professional agents per engine for specific business scenarios.

Govern cost & permissions : Learn billing rules, configure enterprise security and compliance policies.