Visualising Data in Dynatrace

Uploading the Dashboards

This demo comes with several prebuilt dashboards. Do the following in Dynatrace.

• Save the contents of dynatrace/dashboards/openai/Travel-Advisor-Overview.json and dynatrace/dashboards/ollama-pinecone/[AiObs] Travel-Advisor-Overview.json to your computer
• Press Ctrl + k and search for dashboards or select the icon from the left toolbar
• Select the Upload button and upload the JSON file.

Repeat this process for all the dashboards inside dynatrace/dashboards/*

Distributed Traces

The application emits distributed traces which can be viewed in Dynatrace:

• Press ctrl + k search for distributed traces
• Traces for /api/v1/completion are created for each call to either OpenAI or a call to the Weaviate cache.

Remember that only the very first requests for a given destination will go out to OpenAI. So expect many many more cached traces than "live" traces.

Trace of the RAG Pipeline

Tracing LangChain allows to see all step that the pipeline takes to supply external knowledge to the LLM model. In the span attributes, we can observe the prompt crafted by the RAG pipeline with the external documents.

Trace with OpenAI

A "full" call to OpenAI looks like this. Notice the long call halfway through the trace to openai.chat. These traces take much longer (3 seconds vs. 500ms).

Trace to Weaviate Cache

A call which instead only hits the on-cluster Weaviate cache looks like this.

Notice that it is much quicker.

The response TTL (max time that a cached prompt is considered "fresh") is checked and if the response is "still fresh" (ie. TTL < stale time) the cached value is returned.

Notice the cached prompt is 123s. The max age (TTL is (by default) 60 minutes. Therefore the prompt is not outdated and thus returned to the user as valid.

🎉 Demo Complete

The demo is now complete. Continue to cleanup your environment.

• Cleanup resources to avoid GitHub charges