Data Agent deploys autonomous AI agents across your entire data stack — querying, interpreting, and surfacing insights that would take your team weeks to find.
Trusted by 8,000+ data teams worldwide
上周各门店的销售排名是什么？为什么有差异？
Active agents
84,200+
+12% this month
Queries / day
2.4M
+28% vs last week
Avg. latency
1.8 s
−64% faster
Data connectors
120+
plug & play
30+
Agent Nodes
4
Memory Layers
7
Route Decisions
7
Security Layers
Agent Architecture
Each query flows through a deterministic orchestration graph — understanding, planning, executing, validating, and presenting results in a single coherent pass.
Supervisor
4-layer orchestration entry
Query Understanding
Intent + depth detection
Router
7-route intelligent dispatch
Planner
ML DAG decomposition
Data Retrieval
R/V/G hybrid search
Analytical Agent
ML tool execution
SQL Generation
Code-specialized LLM
Code Interpreter
Sandboxed Python
Guardrails
7-layer security validation
Visualization
Auto chart + ML charts
Insights
ML-grounded insights
Suggested Followup
Next-best-action
Storage Architecture
One query. Three synchronized stores. Relational precision, vector recall, and graph governance — working in concert.
R · Relational
Structured DDL, metric definitions, join rules, lifecycle tracking, and ML tool registry.
V · Vector
7 embedding types with pgvector — cosine similarity search for fuzzy term matching and SQL cache.
G · Graph
Apache AGE property graph enforces governance boundaries, relationship traversal, and compliance packs.
Data Governance
Every data asset is governed from definition to query — schema, terminology, and compliance fused into one coherent graph.
Table schemas, column definitions, metric contracts, join rules, and few-shot NL2SQL examples — version-controlled as YAML, ingested into PostgreSQL.
Business term lifecycle management, bilingual disambiguation, synonym networks, and ambiguity matrices — powered by pgvector embeddings.
Compliance rules, business context assumptions, industry archetypes, and regulatory constraints — enforced via Apache AGE graph traversal.
Ingestion Pipeline
Compliance Coverage
GDPR
Enforced
HIPAA
Enforced
PCI-DSS
Enforced
SOX
Enforced
Query Planning Layer
Deterministic graph routing + algebraic reasoning sits between RAG and SQL generation, providing structured guidance so LLMs generate safer, more accurate SQL in complex scenarios.
Pipeline Position
Dijkstra + Kruskal MST finds minimal-cost table join paths with fanout risk weighting.
Three rewrite rules detect and suggest fixes for chasm traps, redundant joins, and semi-join patterns.
Never blocks the pipeline — failures return empty guidance, LLM proceeds independently.
# Query: "Compare Q2 revenue by product across regions with active customer counts"
table: dim_product
No columns referenced, N:1 leaf node
table: fact_orders, fact_sessions
Chasm trap detected: 2 fact tables
table: dim_customer
EXISTS filter detected
Capabilities
Not a wrapper around an LLM. A complete analytical intelligence platform built from first principles.
Four-layer personalized memory (profile, episodic, correction, working) learns your preferences and patterns — every interaction makes the agent smarter.
Every query, span, and decision traced end-to-end. Langfuse for deep tracing, AutoMQ for real-time event streaming.
LLM handles reasoning and planning. Traditional ML handles attribution, forecasting, clustering, anomaly detection, and regression.
What → Why → Next → How. The system detects analytical depth and routes to the right execution path automatically.
Sandboxed Python execution with whitelisted ML libraries. LLM generates analysis code, sandbox executes safely with memory and time limits.
Results as decision-ready blocks — narrative, metrics, charts, ML results, insights, and next-best-action suggestions.
Smart model routing — DeepSeek-v3.2 for understanding, qwen3-coder for SQL, Qwen3-max for insights. Automatic fallback chains.
Convert conversational analysis into reusable BI dashboard widgets. Persist chart configs, SQL, filters, and ML context as durable assets.
Memory Architecture
Unlike stateless LLMs, Data Agent builds a persistent understanding of your team, data, and preferences — becoming more accurate with every query.
Persistent user preferences, domain knowledge, analytical style, and role context. Loaded at session start.
user_role: data_analyst · preferred_chart: bar · domain: fintech
Successful query–result pairs and effective analysis patterns from past sessions. Retrieved via vector similarity.
episode_id: q-3847 · pattern: attribution_model · reuse_score: 0.94
Errors caught, user corrections applied, negative feedback reinforced. Prevents repeating the same mistakes.
correction: avoid_yoy_comparison · trigger: fiscal_year_mismatch
Active session context, in-flight query state, partial results, and real-time clarification thread.
session: s-92f · active_query: revenue_attribution · turns: 6
Observability
No black boxes. Data Agent's ObservabilityFacade instruments every span — from intent detection to chart rendering — with full latency, token, and cost attribution.
Langfuse deep tracing
Full LLM call trees with prompt/response capture and latency breakdown
AutoMQ event streaming
Real-time pipeline events streamed for monitoring and replay
Cost attribution
Token spend, model routing decisions, and cache hit rates per query
» Compare Q2 APAC revenue vs forecast by segment
Security
Defense in depth from authentication to sandbox execution. Every query traverses every layer — no shortcuts, no bypasses.
JWT Authentication
Role-based access control + rate limiting
Permission Filter
Domain-scoped data access boundaries
Metadata Safety
PII exposure policy — hidden / masked / visible
Graph Constraints
Apache AGE enforces governance joins
SQL Policy Engine
SELECT-only whitelist + column validation
Execution Isolation
Dedicated workgroup + statement timeout
Sandbox & Result Masking
Code interpreter whitelist + column-level PII masking
Get Started
Join 100+ data teams who ship insights in seconds, not weeks.