See which model call or tool dominates the request
Use TraceQL to isolate slow model invocations, agent steps and downstream service calls in detailed traces
MODEL TO INFRASTRUCTURE
AI and LLM monitoring
Without a telemetry silo
Store OpenTelemetry spans, metrics and logs from model calls and agent workflows beside the services, databases and infrastructure they depend on. Query the result on a ClickHouse® database backend
ai-llm.observabilityOTEL IN · OPEN QUERIES OUT
AI APPLICATION GenAI instrumentationAgent frameworksModel gatewaysApplication services
STATELESS PULSUSDB
Ingest APIsnormalise · batch · route
Query compilersPromQL · LogQL · TraceQL
STATEFUL DATABASE ClickHouse® database signal-aware schemas
rollups · tiering · scale
rollups · tiering · scale
Complex work is pushed down to the dataReduced results return
01Model call traces02Token metrics03Agent logs04Service telemetry
What this lets you investigate
Keep the operational question
Close to the data
PulsusDB keeps the collection and query surfaces open. The ClickHouse database handles high-volume storage and analytical execution while the signal-specific query language preserves the workflow your team already knows
Track tokens and model traffic with metrics
Aggregate OpenTelemetry GenAI measurements by model, provider, operation and service using PromQL
Keep agent errors connected to the real stack
Search structured logs and traces from the agent, model gateway, vector store and application service in the same platform
Three signals Three query languages
Ask the question
In its native language
PulsusDB compiles familiar observability queries into efficient database work, pushes filters and aggregation toward the ClickHouse database tier and returns only the result needed by the client
PromQL01 / TOKEN RATE
sum by (gen_ai_request_model) (
rate(gen_ai_client_token_usage_total[5m])
)Compare token volume by requested model using standard metric queries
LogQL02 / TOOL FAILURES
{service_name="support-agent",
deployment_environment="production"}
| json
| event="tool_error"Search structured agent logs without sending them to a separate AI-specific store
TraceQL03 / SLOW MODEL CALLS
{ span.gen_ai.system = "openai"
&& duration > 2s }Find slow model invocations and retain the surrounding application trace
One architecture underneath
Optimised for the signal
Unified at the platform
Metrics, logs, traces and profiles do not have the same read path. PulsusDB gives each one a purpose-built physical model, then exposes a consistent OpenTelemetry ingest layer and familiar query APIs above it
Explore the architecture →Signal-specific models
Ordering keys, indexes, codecs and projections follow the dominant query shape for each signal
Stateless application tier
Scale ingest and query processes independently while durable state remains in the database tier
Efficient long retention
Compression, materialised rollups and object storage tiering reduce the cost of detailed operational history
Open client workflows
Use OpenTelemetry for ingest and PromQL, LogQL or TraceQL for investigation
OPEN SOURCE · AGPLV3Bring the telemetry
Bring the telemetry
Keep the query language
Start with the workload in front of you, then add the rest of the estate without adding another observability backend