What This Category of Data Looks Like
Data for this category mainly comes from incident timelines, operation logs, monitoring metrics, and code commit records. Daily output documents include root cause analysis (RCA) reports, issue tracking tickets, fault review minutes, and rectification tracking sheets. Data updates are triggered once per major incident. Monthly centralized archiving and updates are completed. Most data structures combine structured fields and semi-structured text. Fields include incident occurrence time, impact scope, troubleshooting steps, root cause conclusions, rectification measures, and others. There is no unified fixed unit system. Some fields are associated with specific business indicator values.
What Constraints Do These Characteristics Impose on the Vector Model and Indexing Link
Incident-related data updates occur suddenly and in phases. The volume of data updated in a single batch fluctuates based on incident scale. This requires the indexing system to support on-demand incremental updates without full reindexing. Data includes structured fields and semi-structured text. Mixed text types require vector models to adapt to content encoding for different formats. Monthly archived historical data volume accumulates over time. Index storage capacity must accommodate long-term data accumulation requirements. Text related to root cause analysis usually contains professional operation and maintenance terms and technical details. Vector models must have the ability to encode domain-specific text to avoid semantic deviation. Issue tracking tickets and rectification tracking sheets have strong field correlations. The index must support matching logic for multi-field joint retrieval. This document applies to FastGPT v4.17.0.
How to Set Configurations
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of RCA reports and review minutes, avoids including too many irrelevant incident troubleshooting details in a single chunk |
| Overlap Length | 100–200 characters | Retains contextual association between adjacent chunks, prevents key logic of root cause analysis from being truncated by chunking |
| maxToken | 4096 | Adapts to the input length limit of most open-source embedding models, avoids text truncation that loses key information of root cause conclusions |
| Similarity Threshold | 0.75–0.85 | Filters low-match incident logs and documents, focuses on highly relevant root cause troubleshooting evidence |
| Single Search Limit | Top 8 entries | Covers associated troubleshooting records for common incidents, avoids excessive redundant results interfering with root cause localization |
| defaultToken | 1024 | Controls the context output length of a single round of Q&A, ensures clear and complete root cause analysis conclusions |
Three Common Mistakes
- Symptom: A locally deployed M3E model fails to load normally, with an error indicating invalid model path. Cause: The Docker mapping port and absolute path of the model file for the local model are not correctly filled in the FastGPT configuration page, causing the system to fail to access the local embedding model.
- Symptom: When adding the Qwen/Qwen3-Embedding-8B model repeatedly, the original configuration is overwritten. Cause: FastGPT uses model names as unique identifiers, and does not provide configuration options for renaming or multiple instances of the same model.
- Symptom: Field misalignment occurs after chunking of issue tickets and rectification tracking sheets imported from Excel. Cause: The structured data chunking adaptation option is not enabled. The default text chunking logic does not recognize the correspondence between Excel headers and data rows, leading to damaged data integrity during chunking.
How to Confirm Configurations Are Correctly Set
- Upload a small sample RCA report, perform vector index construction, and check whether the text segmentation in the chunk preview interface meets expectations.
- Verify the connectivity of the local embedding model on the model configuration page, trigger a test embedding request, and confirm that the returned vector values are normal.
- Import a batch of Excel issue ticket data that includes headers, and check whether the chunked content retains the association between headers and corresponding data rows.
- Initiate a query related to root cause localization, and check whether the number and matching degree of retrieved results conform to the preset configuration logic.
The values provided here 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.