Databuff
GitHub
Skip to content

中文  |  English

Architecture · AI Platform

Read Telemetry Pipeline and Storage first.

Design Intent

AI in traditional APM is often a "bolt-on chat box" — it cannot read real data and guesses answers.

DataBuff's AI platform was AI-native from day one: AI reads OpenTelemetry APM data directly; multiple experts work together.


Core Principles

PrincipleDescription
Data-drivenAnswers must come from real APM data — no fabrication
Expert specializationDifferent scenarios handled by different experts — more accurate than one model
Brain orchestrationUsers see one entry; complex collaboration runs in the background
Open extensionSkills define behavior; Tools extend capability boundaries

Multi-agent Architecture

Why multi-agent instead of one large model?

Single large modelMulti-agent
AccuracyGeneric answers, easy to stay shallowExperts focus on domain; queries are precise
ExtensibilityNew capability = more prompt, hard to maintainNew expert = new module, isolated
Complex tasksEasy to miss stepsBrain decomposes, experts parallelize, results merge
TrustMay hallucinate dataEach expert must call tools for real data

Three-layer Capability Model

LayerRoleExamples
ExpertUser-facing intelligent roleAI Brain, Query Expert, Inspection Expert
ToolAtomic capabilities experts can callList services, query traces, plot trends
SkillRules constraining expert behaviorQuery semantics, inspection flow, routing

New capability = combine Tools + write Skill + register Expert — no core code changes.


Native Integration with APM

AI is not a separate system — it grows directly on APM data. That means:

  • Ask "error rate" → queries real Doris metrics, not hallucinations
  • Ask "slow traces" → pulls real trace data
  • Ask "service relationships" → draws real topology

This is the essential difference between "AI-native OpenTelemetry APM" and "APM + chat box".


Open Ecosystem

  • Multi-model support: OpenAI-compatible, Anthropic, and other mainstream LLMs
  • MCP protocol: External agents (e.g. Cursor, Claude) can call platform APM tools via standard /mcp; the platform can also act as MCP client to external MCP services
  • Customizable Skills: Built-in skills can be overridden; extend for business scenarios

For external agents connecting to DataBuff as MCP Server, see User Guide · Agent Integration. For external MCP inside the platform and custom Brain-routable experts, see Custom Digital Experts and External MCP Integration.

Built-in AI chat (AgentScope) still uses JAVA_BEAN tool registration; MCP Server is a parallel external path and does not replace platform UI expert binding.