Intelligent operations
Platform capabilities
Security analytics
AI
Dedicated compute
Dedicated compute keeps concurrent queries stable and complex analysis fast.
Learn moreDedicated compute
Dedicated compute keeps concurrent queries stable and complex analysis fast.
Learn moreContact us
Join the community
Try Guance
Start online with usage-based pricing and a true cloud service.
Get startedChoose a Guance edition
The observability of LLMs is not just about whether the model interface is successful, but about explaining the relationships among prompts, model calls, toolchains, vector retrieval, networks, and business logic. Guance LLM observability and large model monitoring solution is based on open capabilities like OpenTelemetry to collect calls, tokens, latency, errors, traces, and logs, helping teams clearly see the entire process from user requests to model responses in a single session.
Model call process black box:A single response may go through retrieval, tool calls, model reasoning, and business services, making it difficult for traditional APMs to explain what happens inside LLM requests.
Token cost and latency are difficult to control:Models, prompts, context length, and call count all affect cost and response time, and without fine-grained data, optimization is difficult.
Errors and quality issues are hard to replicate:Timeouts, rate limiting, null responses, exception outputs, and user feedback need to be analyzed together with the request context, prompt, and model version.
Disconnect between AI applications and business systems:LLM calls are only a segment of the business chain and must be reviewed together with user requests, services, logs, databases, and other upstream and downstream processes.
LLM Calls for Visualization:Record models, requests, prompts, tokens, time, state, and errors to help teams understand call volume, performance, and cost trends.
Trace Link and Flame Diagram Analysis:Model calls, vector retrieval, tool calls, and business services are all grouped under the same trace to identify slow requests and failure stages.
Cost and anomaly alerts:Monitor token consumption, call volume, error rate, latency, and model dimensions to avoid escalating abnormal costs and experience issues.
Open Standard Access:Integrating LLM application data based on the OpenTelemetry ecosystem reduces the cost of integration with existing observable systems.
You will see a real-time updated dashboard clearly showing the total consumed Total Tokens and the ratio of Prompt Tokens to Completion Tokens, and break down average Input/Output Tokens by model and application dimension. With this panoramic monitoring, the team can immediately assess the cost and complexity of each call during development, avoiding sudden bill spikes caused by high consumption after launch; The system also supports setting thresholds for abnormal token fluctuations, triggering an alert immediately if a request exceeds the expected range, helping you stay on top of cost trends at any time.
Book a demo
All LLM API call counts, average cost per call, and cumulative consumption can be aggregated into a single chart. With trend lines, you can gain clear insights into call volume and cost trends, helping product and finance teams plan budgets precisely; At the same time, when a certain dimension's cost or request volume reaches a set threshold, the platform will immediately alert you to prevent potential traffic spikes from overspending budgets, safeguarding your large-scale deployment.
Book a demo
In the "LLM Call Journey History," you can quickly trace the complete metadata of each inference request using Trace ID, including start and end times, the model version actually selected, the temperature parameter, prompt length, and output summaries. This feature not only allows one-click filtering of failed or timeout calls, but also integrates input and output logs for deep troubleshooting, helping development and operations teams quickly reproduce issues; It also supports exporting historical records in CSV or JSON formats, facilitating cross-departmental sharing and analysis.
Book a demo
The "Flame Graph and Root Cause Analysis" module automatically renders the time consumption distribution of sub-Span on the link detail page, allowing you to intuitively identify which call phases are the most time-consuming, and by combining response time curves such as P75, P90, and P99, you can deeply explore the root causes of occasional slow requests. Once you identify a bottleneck, you can instantly jump to the corresponding code or configuration interface, helping your team complete performance diagnostics and optimization within the same platform.
Book a demo
DevOps is observable
Integrate R&D, release, operation, and alarm response to improve delivery quality.
Log management and analytics
Centralized retrieval, parsing, and correlation of multi-source logs to pinpoint the root causes of anomalies.
Observability platform and unified monitoring platform
Unified metrics, logs, links, RUM, AI analytics, and infrastructure data.
Pay attention to model call volume, token consumption, response time, error rate, timeout, model version, prompt context, trace chain, and business impact.
You can move from call trends to single-time traces, viewing models, prompts, tokens, tool calls, retrieval, and service time spent to identify bottlenecks.
OpenTelemetry-related capabilities can collect calls, links, logs, and metrics from LLM applications, and analyze them in a unified manner with existing service monitoring.
Let Guance match your LLM observability with the implementation path of large model monitoring solutions
Schedule LLM observation demonstrations
