Cisco WebexAI service diagnostics

From Unified Real-Time Analytics to a Governed AI Data Plane

Cisco Webex moved its analytics platform from Apache Pinot, Trino, TiDB and Spark/Kyuubi to one StarRocks data plane, then built the Webex AI Assistant on top of it. Support and engineering teams ask about a meeting, call or device in plain language, and the assistant answers with SQL that StarRocks runs against live service data.

3

product areas the AI assistant diagnoses

Meetings, Calling and Devices

6

data centers served

Regional deployments; Japan and London planned

CompanyCisco Webex
Use caseAI service diagnostics
Previous solutionApache Pinot, Trino, TiDB and Spark

This story is based on a NextQuery 2026 session by Hongkun Xu, Senior Software Engineer at Cisco. NextQuery is PhoenixAI's Data + AI Infrastructure Summit. Watch the session

About Cisco Webex

Webex is Cisco's enterprise collaboration platform. It provides meetings, calling, messaging and contact center services to businesses worldwide, connecting employees, customers and devices.

Every Webex interaction produces live business and operational signals. Meetings produce sessions, quality data and transcripts. Calling produces call records and media quality. Messaging, the contact center and devices add activity, interactions, telemetry and health data.

These signals are related, so a useful answer rarely sits in one source. A question about a poor meeting can involve the client, the network, the device and the service. The data also spans six deployed data centers that support regional workloads, and a single data center can run several clusters.

The challenge

Webex wanted AI applications that could answer questions from its meeting, calling and device signals. The data path at the time made that hard. Queries ran across Apache Pinot, Trino, TiDB and Spark with Kyuubi, over data in S3 or HDFS. Duplicated engines, inconsistent semantics and separate access paths made fresh, governed AI hard to build.

Four blockers stood in the way:

  • Fragmented access paths. Engines and copies of the data multiplied.
  • Inconsistent metrics and schemas. Definitions diverged between systems.
  • Stale retrieval context. Offline indexes drifted from the facts they described.
  • Manual diagnosis. Troubleshooting evidence lived with a few experts.

Each of these gets worse when an AI agent is the one asking. An agent that has to pick between copies of the data, or reconcile two definitions of the same metric, gives answers no one can trust.

Diagram of the move from fragmented engines to one analytical plane: Pinot, Trino, TiDB and Spark/Kyuubi, with S3/HDFS storage, all point to a single StarRocks box offering unified SQL, real-time plus lakehouse access, shared semantics and one governed access path. A banner reads: remove fragmentation before adding an AI interaction layer
From fragmented engines to one StarRocks analytical plane. Source: Cisco Webex, NextQuery 2026

The solution

One data plane for every consumer

Cisco Webex first consolidated its real-time analytics on StarRocks, as told in How Cisco Webex Unified Pinot and Trino Workloads on StarRocks. The wider platform moves from Pinot, Trino, TiDB, Spark/Kyuubi and S3/HDFS to one StarRocks analytical plane, which Webex put in place before adding any AI layer.

Webex data now lands once in StarRocks, through Kafka, ETL jobs and lakehouse sources, and becomes queryable and retrievable in one place.

The StarRocks data plane holds real-time tables, lakehouse catalogs, JSON data, materialized views, a full-text index and a vector index. RBAC with Apache Ranger, auditing, resource groups and action controls govern every access path. The design principle is one copy of truth and one retrieval and query plane.

Four consumers read from it: BI dashboards built on known SQL and metrics, product services through low-latency data APIs, business Q&A through the AI assistant, and evidence-guided troubleshooting. Because they share one copy, a number in a dashboard and a number in an AI answer come from the same place.

Reference architecture: Webex Meetings, Calling, Contact Center and Messaging data flows through Kafka, unified ETL jobs including Apache Spark, and lakehouse sources into one StarRocks data plane with real-time tables, lakehouse catalogs, materialized views, JSON, full-text and vector indexes, under RBAC, audit, resource-group and action-control governance. The plane serves BI and dashboards, product services, the Webex AI Assistant and troubleshooting. Below, the assistant's five steps: resolve intent, generate SQL, query StarRocks materialized views first, generate the answer, generate a chart
Cisco Webex's StarRocks data plane and the Webex AI Assistant

For AI agents, Webex needed fresh facts, fast evidence and a traceable execution surface, rather than another offline copy of the data. Each maps to something StarRocks provides:

  • Fresh. Streaming ingestion, direct lakehouse access and update-capable tables keep answers aligned with the current state of Webex.
  • Fast. MPP execution, vectorized SQL, materialized views and indexes return evidence at conversational speed.
  • Verifiable. SQL and tool traces, audit logs, identities and resource controls let every answer show its work.

The Webex AI Assistant

The Webex AI Assistant is a diagnostic service for Webex Meetings, Calling and Devices. It combines Webex troubleshooting knowledge with live service evidence: quality, join failures, client errors and ThousandEyes analysis for meetings; call lists, media quality and SIP call diagrams for calling; and join timing, failures and client errors for devices.

Users start from what they have, such as an email address, conference ID, meeting number, phone number, call ID or correlation ID, plus a time window. The assistant sits next to the live operational view, with sample prompts for each area.

Under the hood it has three layers. Kafka and unified ETL and Apache Spark jobs bring in Webex data. StarRocks handles storage and analytics, with OLAP tables, materialized views, a vector store and a text index. The AI layer, built with LangChain, OpenAI GPT-5 and OpenAI embeddings, handles aggregation, on-demand queries, vector retrieval, text queries and chart generation for troubleshooting, trend analysis and Q&A.

The Webex AI Assistant architecture in three columns. Data ingestion and processing: Webex Meetings, Calling and Contact Center feed Kafka and unified ETL and Apache Spark jobs. Data storage and analytics: StarRocks with OLAP, materialized views, a vector store and a text index. AI-powered business features: aggregation, on-demand query, vector retrieval, text query and chart generation around AI tools (LangChain, OpenAI GPT-5, OpenAI Embedding), with history context, Text-to-SQL, troubleshooting, trending analysis and Q&A
Webex AI Assistant architecture. Source: Cisco Webex, NextQuery 2026

Accurate answers rest on a few elements. StarRocks materialized views supply comprehensive, high-quality diagnostic data. Retrieval-augmented generation adds a domain knowledge base. Prompt engineering and chain-of-thought reasoning handle intent recognition and schema-based SQL generation. Personal data is encrypted or masked before anything reaches the model.

How the Webex AI Assistant answers a question

A Cisco-proxied OpenAI model handles the assistant's language steps, and StarRocks does the data work:

  1. The assistant reads the question and the conversation so far to understand the intent and resolve ambiguity.
  2. It generates SQL from the relevant schemas and prompts, choosing the business tables, columns and conditions it needs.
  3. StarRocks runs the query, using materialized views built on pre-aggregated business data first so the answer comes back quickly.
  4. The assistant combines the question with the result to write a precise, readable answer.
  5. When a comparison or pattern helps, it draws a chart.

Users do not need to understand the underlying data tables or move across multiple dashboards.

In one meeting quality review, a Webex engineer asked the assistant to review a user's meeting quality over the past seven days and check for a network issue. The assistant pulled meeting history, media quality and client distribution.

It found audio and video consistently good. Content sharing showed a brief latency spike in one meeting, which did not make it a poor session. Because both meetings ran on Wi-Fi, the assistant attributed the spike to momentary network fluctuation rather than a chronic server-region issue. Charts of client versions and IP addresses showed any recurring pattern.

The engineer did not need to open several dashboards or ask a platform expert, and the answer pointed to the next thing to check.

Evidence-first troubleshooting for the Webex data platform

Cisco Webex applies the evidence-first approach behind the Webex AI Assistant to running its StarRocks data platform. Reusable diagnostic skills triage a symptom, query load state, logs, metrics and query profiles, rank likely causes with confidence, and recommend the smallest safe fix before verifying the outcome. Read-only actions are the default, reversible actions run automatically only when safe, controlled changes need approval and a rollback path, and disruptive actions are led by experts.

Write failures and slow queries share one evidence spine of time, scope, transaction IDs, profiles, logs and resource state. For a write failure, the troubleshooting assistant follows the load record through the transaction and publish step to the FE and BE logs. In an illustrative case, it traces a spike to a commit bottleneck on one node and proposes isolating the node and retrying a small batch. For a slow query, it starts from the query profile, finds the hottest operator and explains why a fix matches that operator's signature.

StarRocksApache KafkaApache SparkApache RangerLangChainOpenAI GPT-5

The results

What one data plane gives Webex

With one data plane in place, Webex now serves dashboards, product services, AI answers and operations from the same governed data:

  • One copy of the data. Webex data lands once in StarRocks and serves BI, product services, business Q&A and troubleshooting through one governed access path.
  • Diagnosis from a single question. Support and engineering users describe a meeting, call or device problem in plain language and get an analysis back in the conversation after a short wait, without knowing the tables or switching dashboards.
  • Answers that show their work. Each answer is built from SQL that StarRocks runs on current data, with materialized views keeping the response fast.
  • One method across business and operations. The evidence-first approach behind the assistant also structures how Webex handles write failures and slow queries.

What's next

Three principles guide Webex's work ahead: unify Webex live signals into governed facts, ground every AI answer in current context and query evidence, and act with controls, recommending the smallest safe intervention and verifying the outcome.

Webex has data centers planned in Japan and London, and FedRAMP runs as a separate regulated environment. On the data plane, Webex is validating full-text search and the vector index, both Beta in StarRocks v4.1, for broader use. Both already back the Webex AI Assistant's vector retrieval and text queries, so wider use keeps its answers on the same governed copy of Webex data that dashboards and operations tools read.