What this category’s data looks like
This category’s data is collected from DCS/SIS systems, online analyzers, LIMS systems, equipment condition monitoring modules, energy metering systems, and raw material storage tank ledgers. Daily documents include operating procedures, process index cards, timestamped DCS trend charts, laboratory reports, interlock bypass records, special work permits, and shift logs. Data update frequencies cover real-time (such as DCS trends, online analyzer readings), per-shift (such as shift logs, special work permits), and daily (such as monthly laboratory summary reports). Document fields include process parameters (such as reaction temperature and pressure, with units of ℃ and MPa), equipment numbers, interlock logic, sampling times, and qualified judgment results. The data structure includes both structured numerical reports and unstructured text records.
How these characteristics impose constraints on knowledge base retrieval and recall
Multiple heterogeneous data types require the retrieval system to support both structured field matching and full-text search. This prevents missing key professional information such as process parameters and interlock numbers. Documents with different update frequencies must be distinguished by timeliness. Real-time data must recall the latest versions first. Outdated old operating procedures must be filtered out. The inspection and audit scenario requires accurate matching of multiple related documents. A single retrieval result must cover process requirements, real-time data, and historical records. It must avoid retrieving irrelevant non-process documents. Incorrect retrieval results may cause production accidents. Low-match irrelevant content must be strictly controlled from entering the context.
How to configure the settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Basic chemical process parameters have clear thresholds. This range filters low-match irrelevant documents and avoids mistakenly recalling outdated incorrect process indicators |
limit | Top 6 entries | Inspection and audit require checking multiple related documents. This quantity covers core content such as process index cards, DCS trends, and interlock records, and does not exceed context window limits |
usingReRank | Enabled | Re-sorts initial retrieval results by professional relevance, prioritizes returning key audit basis such as interlock records and process requirements |
datasetSearchUsingExtensionQuery | Enabled | Performs extended decomposition on the user's inspection and audit questions, matches multiple related documents, and avoids retrieving only isolated single entries |
searchMode | fullTextRecall | Use full-text retrieval for text content. Field, metadata, batch, and document-type constraints require separate filters and are not enforced by this mode. |
Three common configuration errors
- Phenomenon: The retrieval result only returns a single citation link, with no corresponding answer content. Cause:
datasetSearchUsingExtensionQueryis not enabled. The user's inspection and audit questions are not extended and decomposed. The system cannot match multiple related process documents and records. - Phenomenon: Structured fields such as process parameters and interlock numbers cannot be correctly retrieved from imported docx and excel documents. Cause: The retrieval mode for matching structured fields is not configured. Relying only on full-text search cannot accurately match professional field content.
- Phenomenon: After importing Feishu documents, some text descriptions attached to images cannot be retrieved. Cause: The image text extraction function is not enabled, or the nested format of Feishu documents is not adapted. This causes some content to not be correctly indexed.
How to verify the configuration is complete
- Upload a test process index card and DCS trend document. Initiate the test question "Check whether the current reactor pressure meets process requirements" in the FastGPT v4.17.0 environment. Check whether the retrieval result includes the threshold and real-time value of the corresponding document.
- View the knowledge base index log. Confirm that all imported document types (docx, excel, Feishu documents) have generated index entries, with no format errors.
- Adjust the
similarityparameter value. Test the number of retrieval results under different thresholds, and confirm that the filtering effect meets the needs of the professional scenario. - After enabling
usingReRank, test whether the sorting of retrieval results for the same question better aligns with audit priorities, and prioritizes returning interlock records and process requirement documents.
The values provided here are common starting points and should be measured against the reader's own samples.