What the data for this category looks like
The primary data for this category comes from production-related documents of basic chemical enterprises including chlor-alkali, synthetic ammonia, methanol, and soda ash. Daily documents cover operating procedures, process index cards, DCS trends, laboratory reports, interlock bypass records, special work permits, and shift logs. Data sources include DCS/SIS systems, online analyzers, LIMS systems, equipment condition monitoring, energy metering systems, and raw material tank data. Update frequencies are split into three categories: real-time (such as DCS trends), per-shift (such as shift logs), and daily (such as laboratory reports). Documents include clear process parameter fields, with temperature units in ℃, pressure units in MPa, and flow units in m³/h. Time-series documents have timestamps, and tabular documents include fixed headers and numerical rows.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The multi-source heterogeneous data, differentiated update frequencies, and high-risk business attributes of this category create multiple constraints for the retrieval and recall process. Real-time updated DCS trends and shift log data require an incremental synchronization mechanism to keep the knowledge base content consistent with production site data. Multiple document types cover structured tables, time-series data, and unstructured text, so a hybrid retrieval mode is needed to cover content across different formats. The recall accuracy of documents related to strong interlocks directly relates to production safety, so a high matching threshold must be set to filter low-relevance results. The document system organized by production unit requires limiting the retrieval scope to the corresponding workshop or process line, to avoid irrelevant cross-unit content interfering with review work.
How to set the configurations
| Configuration Item | Recommended Value | Rationale for this Setting |
|---|---|---|
similarity | 0.75–0.85 | Basic chemical documents contain strong interlock parameters. Filter low-match results to avoid incorrectly recalling irrelevant process content |
limit | Top 8 entries | Single-batch recall volume aligns with the short-text characteristics of shift logs and process cards, preventing excessively long context that reduces review efficiency |
searchMode | mixedRecall | Supports hybrid retrieval for structured data (DCS trends, laboratory reports) and unstructured text, covering all supported document types |
usingReRank | Enabled | For high-risk documents such as interlock bypass records, re-ranking improves the sorting priority of key content |
datasetSearchUsingExtensionQuery | Enabled | Expands queries using process terminology to cover professional abbreviations in DCS/SIS systems, such as "PV" corresponding to "process variable" |
collectionIds | Divided by workshop/process line | Matches the production-unit-based document system of basic chemical enterprises, narrows the retrieval scope, and improves recall accuracy |
This configuration applies to FastGPT v4.17.0.
Three common configuration mistakes
- Empty retrieval results after uploading a process index card file. The cause is that the corresponding document was not added to the specified
collectionIdscollection, so the retrieval scope does not cover the target document. - A 403 status code is returned when calling the knowledge base content addition interface. The cause is that reverse proxy permissions for intranet access were not configured, blocking the interface call.
- Only table headers are returned during retrieval of tabular data such as laboratory reports, with no specific numerical values. The cause is that table multi-vector support was not enabled, so table cell content cannot be split for recall.
How to confirm the configuration is properly set
- Upload an interlock bypass record from the corresponding workshop, and verify that the retrieval results include key interlock numbers and disposal steps.
- Navigate to the FastGPT knowledge base settings page, and confirm that the table multi-vector support function is enabled.
- Access the intranet address
https://domain:3000, and verify that files can be uploaded normally and the knowledge base addition interface can be called successfully. - Check the database storage logs, and confirm that trained documents have been written to the knowledge base table for the corresponding tenant.
The values provided are common starting points and should be measured against the reader's own samples.