What the Data for This Category Looks Like
Typical data for this category comes from highway maintenance management systems, bridge and tunnel monitoring systems, road patrol systems, and traffic flow monitoring data. The update frequency is daily. Daily documents include maintenance plans, patrol logs, defect disposal orders, road-related construction permits, road property damage compensation orders, and other types. Document structures cover fields such as patrol point coordinates, defect types, disposal process numbers, permit validity periods, and more. Units include kilometers, square meters, working days, and others. Field formats vary across different documents. Some are structured forms, while others are unstructured log text.
What Constraints These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Heterogeneous data sources from multiple systems require the retrieval pipeline to support cross-data-source document integration. The daily update frequency requires the knowledge base to be configured with an incremental synchronization mechanism to ensure data timeliness. Document types include structured forms and unstructured logs. Keyword and semantic features differ significantly across documents, so a hybrid retrieval mode must be adapted. Core business process links rely on accurate process guidance. The accuracy of recall results directly affects work safety and compliance. Therefore, the retrieval link must avoid interference from irrelevant documents while covering key business documents.
How to Configure Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Balances recall coverage and accuracy, adapts to the semantic matching requirements of highway maintenance documents, and avoids recalling irrelevant patrol logs or construction permits |
limit | Top 5–10 entries | Covers the number of core reference documents for daily business processing, while controlling context length to avoid exceeding the processing limits of large language models |
searchMode | hybrid | Combines keyword and semantic retrieval, adapts to mixed structured and unstructured business documents such as patrol logs and work orders |
usingReRank | Enabled | Improves the ranking accuracy of key process documents such as defect disposal and road-related construction, ensuring core guidance documents are returned first |
datasetSearchUsingExtensionQuery | Enabled | Supplementary extended queries for business terms such as road-related construction and road property damage, improving recall relevance across documents |
collectionIds | Configured by business module | Associating exclusive datasets for maintenance plans, patrol logs, and other content separately, avoiding cross-module document interference with retrieval results |
Three Common Misconfigurations
- Symptom: A "no matching results" error is returned when using the knowledge base, and
text-embedding-ada-002has been registered in OneAPI and added to the configuration file. Cause: The model has not been associated in the vector model configuration of the corresponding dataset, or there are case or spacing differences between the model identifier in the configuration file and the registered name in OneAPI. - Symptom: After local deployment, the content associated with the knowledge base does not appear in the response body, but the citation list shows that the correct documents were matched. Cause: The
usingReRankconfiguration is not enabled, orsearchModeis only set to semantic retrieval, causing key documents to be ranked too low and truncated. - Symptom: After sending a conversation request via the API, the generated response does not reference knowledge base documents. Cause: The
collectionIdsparameter is not specified in the conversation request, or the dataset pointed to by the parameter has not had business documents uploaded and indexed.
How to Confirm Configuration Completion
- The dataset configuration page in FastGPT v4.17.0 may be reviewed to confirm the vector model is
text-embedding-ada-002and matches the registered identifier in OneAPI exactly. - A simulated business query, such as "pavement pothole disposal process", may be submitted. The returned recall results may be checked to confirm they include defect disposal orders and maintenance plan-related documents.
- A conversation request via the API may be sent to confirm the request parameters include the correct
collectionIdsandsearchModeconfigurations. - The retrieval log may be reviewed to confirm the number of recalled documents matches the value set in the
limitparameter.
The values provided here are common starting points and should be evaluated against local deployment samples.
Question material comes from public community discussions. Configuration values are common starting points and should be measured against your own samples. Verified on 2026-09-28.