What this category of data looks like
Data for this category primarily comes from tower operation and maintenance management systems in the state-owned enterprise communication sector, daily inspection mobile collection tools, and base station completion acceptance archives. Daily inspection data is updated synchronously each day. Quarterly and annual summary ledgers are updated in batches each month.
Each individual ledger document includes the unique tower ID, installation longitude and latitude, construction completion date, maintenance cycle, previous inspection records, list of attached equipment, and fault repair records. Fields and units include tower height in meters, number of inspections in counts, maintenance cost in yuan, and longitude and latitude use decimal coordinate format.
What constraints do these characteristics impose during the knowledge base retrieval and recall stage?
The data contains both structured identifiers and unstructured text. Retrieval must support both exact matching of IDs and semantic matching of fault descriptions. Individual ledger datasets are large in scale, so the number of single-call retrieved results must be limited to avoid context overflow. There are differences in data update rhythms, so support for switching between incremental and full retrieval is required. Attached equipment and fault records are closely linked, so the relevance of retrieved results must be ensured. Some inspection records include image attachments, so mixed text-image retrieval must be supported to guarantee the completeness of summary reports.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.78 | The accuracy requirements for tower ledger fields are high. This range filters low-match irrelevant inspection records and prevents incorrect data from being included in summary results |
limit | Top 15 entries | A single ledger summary needs to cover inspection records from the past three months. 15 entries cover core data from a typical maintenance cycle and avoids overly long context |
searchMode | vector+keyword | This balances structured fields such as tower ID and longitude/latitude with unstructured fault description text, and hybrid retrieval improves matching accuracy |
usingReRank | Enabled | Performs secondary sorting on initial retrieved results to distinguish the relevance of inspection records and fault repair records, and optimizes the content order of summary reports |
datasetSearchUsingExtensionQuery | Enabled | Automatically supplements unit keywords associated with fields (such as "meters", "yuan") to reduce retrieval bias caused by missing units |
collectionIds | Specify a dedicated tower operation and maintenance knowledge base collection ID | Isolates knowledge base data from other categories, prevents irrelevant content from being included via cross-collection retrieval, and ensures the specificity of summary reports |
Three common mistakes
- Phenomenon: No GPU resource usage during retrieval, with empty returned results or error messages. Cause:
collectionIdsis not specified correctly, and the retrieval scope is not linked to the knowledge base collection containing tower operation and maintenance data, resulting in no valid retrieved data. - Phenomenon: After modifying configuration items and saving, parameters revert to default values after page refresh. Cause: Final saving was not completed on the Knowledge Base Settings page for FastGPT v4.17.0, or browser caching caused configuration to not sync to the server.
- Phenomenon: Retrieved results only contain plain text, with no associated information for inspection photos. Cause: Text-image chunking rules are not configured, or
usingReRankis not enabled to perform associated sorting on text-image metadata, resulting in image attachments not being retrieved correctly.
How to confirm the configuration is complete
- Access the FastGPT Knowledge Base Management page, review the configuration items for the current collection, and confirm that parameters such as
similarityandlimitmatch the preset values. - Submit a retrieval request for a specific tower ID, and verify that the number of returned results matches the top 15 entries set by
limit. - Submit a search term that includes units (such as "tower height 80 meters"), and confirm that the retrieved results include ledger records with the corresponding height.
- Check server logs to confirm there are no retrieval timeouts or parameter parsing errors, and that the GPU monitoring panel shows resource usage for embedding and retrieval tasks.
The values provided are common starting points and should be measured 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 2026-09-28.