What does the data for this category look like?
Data sources include monitoring alerts, log platforms, APM tools, and user feedback. Core documents cover incident tickets, event timelines, emergency response plans, and incident notifications. Data updates occur in real time or after each incident. Fields include incident ID, affected service scope, occurrence time, stop-loss operation steps, associated alert rules, and more. Some structured data uses standardized units such as timestamps and service names. Unstructured documents record processing workflows and review content in natural language.
What constraints do these characteristics impose on the knowledge base retrieval and recall link?
The real-time update nature of the data requires the retrieval link to support incrementally synced datasets. This prevents recall of expired incident handling records. The mixed document structure—structured tickets and unstructured plans—requires retrieval to support mixed recall of multi-format content. It must balance keyword matching and semantic understanding. The high-frequency update feature requires recall results to align with the incident response window, with no delayed returns. Fields associated with incident ID and affected services require retrieval to support precise filtering by specific fields. This narrows the recall scope and improves decision-making efficiency.
How to set up configurations
The following configurations are adapted for FastGPT v4.17.0:
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches core semantics of incident-related documents, filtering out content with low relevance to the current query |
limit | Top 3–5 results | Controls the number of single-recall results, adapting to fast decision-making scenarios in incident response and avoiding information overload |
searchMode | hybrid | Combines keyword and semantic retrieval, adapting to the mixed data characteristics of structured tickets and unstructured emergency plans |
usingReRank | Enabled | Reranks recall results to prioritize returning plans or ticket steps that best match the current incident scenario |
datasetSearchUsingExtensionQuery | Enabled | Generates extended queries based on user questions to cover implicit needs in incident scenarios, such as associated monitoring alert rules |
collectionIds | Dataset IDs bound to the incident scenario | Limits the retrieval scope to the knowledge base collection associated with the current incident, excluding irrelevant dataset content |
Three common configuration errors
- Symptom: The number of single-recall results does not match the configured
limitparameter. Cause: Failed to correctly bindcollectionIdsto limit the retrieval scope, or failed to adjust thelimitvalue to match the dataset capacity. - Symptom: AI responses include content outside the knowledge base. Cause: Did not enable the configuration to only retrieve knowledge base content, leading the model to use external generation logic.
- Symptom: Retrieval results contain a large number of duplicate incident handling steps. Cause: Did not set a reasonable chunk repetition rate filter, resulting in multiple recalls of similar chunks.
How to verify the configuration is correct
- Submit a test question related to the current incident scenario, and confirm the number of returned results matches the
limitsetting. - Check that all content in the AI response comes exclusively from the bound knowledge base collection, with no external unrelated content.
- Review the retrieval interface’s returned results to confirm the format follows JSON specifications, with no unescaped special characters.
- Simulate updating an incident ticket, then submit a related question. Confirm the retrieval results include the latest event records.
The values provided are common starting points and should be evaluated against the reader’s own samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-10-19.