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
| Principle | Description |
|---|---|
| Data-driven | Answers must come from real APM data — no fabrication |
| Expert specialization | Different scenarios handled by different experts — more accurate than one model |
| Brain orchestration | Users see one entry; complex collaboration runs in the background |
| Open extension | Skills define behavior; Tools extend capability boundaries |
Multi-agent Architecture
Why multi-agent instead of one large model?
| Single large model | Multi-agent | |
|---|---|---|
| Accuracy | Generic answers, easy to stay shallow | Experts focus on domain; queries are precise |
| Extensibility | New capability = more prompt, hard to maintain | New expert = new module, isolated |
| Complex tasks | Easy to miss steps | Brain decomposes, experts parallelize, results merge |
| Trust | May hallucinate data | Each expert must call tools for real data |
Three-layer Capability Model
| Layer | Role | Examples |
|---|---|---|
| Expert | User-facing intelligent role | AI Brain, Query Expert, Inspection Expert |
| Tool | Atomic capabilities experts can call | List services, query traces, plot trends |
| Skill | Rules constraining expert behavior | Query 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.