What this category of data includes
The data sources for this category are monitoring alerts, log platforms, APM tools, and user feedback. Daily documents include trouble tickets, incident timelines, emergency response plans, and fault notifications. Fields cover incident ID, impact scope, troubleshooting steps, notification time, and more. Data updates occur in real time or after a single incident is triggered. Document structures combine structured ticket fields and unstructured emergency content, to support rapid retrieval and decision-making in operational on-call scenarios.
Constraints for Knowledge Base Retrieval and Recall
Real-time or event-triggered data updates require retrieval and recall systems to support incremental synchronization and low-latency queries. This prevents outdated fault records from interfering with current decision-making. Documents combine structured fields and unstructured content, so retrieval systems must balance the accuracy of keyword matching and vector retrieval. This ensures core fields such as impact scope and troubleshooting steps are accurately recalled. In fault scenarios, valid content must be located quickly. Retrieval result count and ranking must align with on-call personnel’s quick browsing habits. This avoids redundant information increasing decision-making time. Fault data involves service interruption risks, so retrieval result relevance must be strictly controlled. This prevents unrelated documents from misleading troubleshooting operations.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Fault scenarios require precise matching of troubleshooting steps and impact scope. This range filters low-relevance results to avoid interfering with rapid decision-making |
limit | Top 8 results | Operational on-call scenarios require quick browsing of candidate results. Too many entries increase decision-making time, and 8 results balances recall volume and browsing efficiency |
searchMode | Hybrid vector and keyword retrieval | Daily documents include structured ticket fields and unstructured emergency response plans. Hybrid retrieval balances accuracy and recall completeness |
usingReRank | Enabled | Fault scenarios require prioritizing high-value results that include troubleshooting operations and impact assessment. Reranking optimizes the ranking logic |
datasetSearchUsingExtensionQuery | Enabled | Incident ID and impact scope from fault tickets can be used as extended query terms, improving recall rates for similar fault records |
collectionIds | Bind dedicated collections for fault tickets and emergency response plans | Isolate non-operational knowledge base content to avoid unrelated results interfering with fault response workflows |
Three Common Mistakes
- Symptom: Entries with a full-text search score of 1.2717 appear before entries with a score of 0.5208 in retrieval results. Cause: The
usingReRankfunction is not enabled, and only raw vector similarity ranking is used, without adjusting ranking logic based on business weights for fault scenarios. - Symptom: Importing Excel format files fails, and CSV files only support two columns of data. Cause: Document parsing extension capabilities are not enabled, or field mapping rules for multi-column data are not configured.
- Symptom: Unable to filter retrieval scope by folder. Cause: The retrieval scope is not limited to the specified
collectionIds, or folders are not bound to dedicated knowledge base collections.
How to Verify Configurations
- Access the knowledge base configuration interface of FastGPT v4.17.0, and verify that the values of parameters such as
similarity,limit, andusingReRankmatch the preset configurations. - Upload a test fault ticket document, initiate a retrieval containing impact scope keywords, and confirm that the ranking of returned results aligns with business expectations.
- Attempt to import an Excel format test document, and confirm that multi-column field content can be extracted normally after parsing.
- Select the
collectionIdsbound to the specified folder, initiate a retrieval, and confirm that only knowledge base content within that folder is returned.
The values provided are common starting points and should be evaluated against the user'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.