Between March 31 and April 1, 2026, we experienced an indexing incident that caused some knowledge bases to be incompletely reflected in retrieval. In affected cases, recently updated content or some documents in a knowledge base were not fully available to the system during answer generation, which could cause responses to rely on an incomplete view of the latest knowledge base content.

The issue was caused by a combination of indexing pipeline performance pressure and an error in how the latest knowledge base updates were carried into retrieval. Together, these led some indexing jobs to back up and, in some cases, prevented the most current knowledge base content from being fully available during answer generation. This issue affected retrieval freshness only; source content was not lost.

We have now:

• Applied database and queue mitigations to stabilize indexing throughput

• Corrected the indexing logic so the latest knowledge base updates are now reflected correctly in retrieval

• Rerun affected indexing jobs and validated the results for impacted knowledge bases

• Confirmed the service has returned to normal operation

We are also continuing follow-up work to reduce the likelihood of recurrence, including:

• Stronger isolation between bulk backfills and normal indexing traffic

• Additional safeguards around indexing retries and queue pressure

• Expanded monitoring and alerting for indexing freshness and backlog conditions

• Further improvements to make indexing more efficient and resilient as knowledge bases change over time