Skip to main content
Version: 2.0

Query Observability

As developers build generative AI applications, one of the biggest challenges is understanding why some queries fail to produce the expected results. Diagnosing these issues can be complex, and without clear insights into query execution, users are left guessing about the root cause of problems. This makes it difficult to debug and improve query performance.

Vectara’s query observability provides users with detailed insights into the execution of past queries by offering a comprehensive breakdown of each query’s configuration, response, and execution flow. This allows users to inspect queries at each stage of the query lifecycle which helps diagnose, debug, and optimize their searches.

tip

For more details about query configurations, see Configure Queries.

Query histories​

The Query histories tab presents a table of previous queries executed against a selected corpus. A query history explains the actions taken by our platform in service of the query.

Use this table to inspect query details, mark queries for special consideration, and annotate them with comments. The table lists each query ID, the time of the query, the mode (such as Summary), the query text, query response, latency, and any errors.

Query histories

You can inspect each query by selecting an ID to see how the platform interpreted the query configuration and input, and how the system acted on this input, along with the responses and errors.

Query details​

The query details page provides detailed visibility into how a our platform processed the query, offering valuable insights for both troubleshooting and optimization. These details include the query text, search results, generated response, and factual consistency score.

For example:

Query analysis

This example query detail shows the following configuration for the query:

  • The application type is Summary.
  • The hybrid search parameters set the lambda value to 0.025.
  • Reranking is enabled and using Rerank_Multilingual_v1.
  • The relevance tuning does not show any custom dimensions.
  • The result context is 2 sentences before and after.
  • This query used the Mockingbird 1.0 prompt template with an English summary language and 5 summarized results.
  • The user enabled FCS with both 2 sentences before and after.

Click Expand all to reveal detailed information for each execution stage, including specific parameters.