What this category of data looks like
Pulp and paper inspection and process record data sources include pulp and waste paper batch ledgers, DCS/QCS control system data, on-machine measurement data, laboratory quality inspection reports, energy and water treatment system records, and sheet break and grade change ledgers. Data update cycles cover three levels: real-time, per shift, and per roll. Document types include pulp preparation sheets, beating records, paper machine operation logs, online quality curves, laboratory test reports, and sheet break and grade change records. Document fields include pulp ratio, beating degree, grammage, moisture content, and thickness. Corresponding units are percentage, °SR, grams per square meter, percentage, and millimeters.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Multi-source heterogeneous data types and layered update cycles require the retrieval system to support matching both text and numeric fields, and adapt to real-time incremental update requirements. Process record entries for a single paper roll are numerous. The number of returned search results must be controlled within a reasonable range to avoid exceeding the context window. Documents are clearly categorized by production stage. Retrieval must support limiting scope by production department to reduce irrelevant result interference. The specificity of professional terminology and industry parameters requires the retrieval model to adapt to pulp and paper-specific vocabulary to avoid semantic matching deviations.
How to set the configurations
This guide applies to FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.65–0.75 | Normal process fluctuations exist in pulp and paper quality data. An excessively high threshold will filter critical parameter records, while an excessively low threshold will introduce a large number of irrelevant results. |
limit | Top 8–12 entries | Process record entries for a single paper roll are numerous. Returning too many results will exceed the context window and reduce subsequent analysis efficiency. |
searchMode | mixedRecall | Balances retrieval needs for text-based documents such as pulp preparation sheets and logs, and numeric-based documents such as online quality curves and quality inspection reports. |
usingReRank | Enabled | Fields in process records have strong correlation. Reranking can filter historical batch data irrelevant to the current query and improve result relevance. |
datasetSearchUsingExtensionQuery | Calibrate based on actual measurements | Different factories have different technical terminology habits. Optimized queries can improve matching accuracy for professional terms. |
collectionIds | 3–5 department collection IDs | Documents are clearly categorized by department. Limiting collections can narrow the retrieval scope and improve retrieval speed. |
Three Common Mistakes
- After uploading knowledge base documents in an offline environment, the interface displays a loading spinner with no status change for an extended period. Cause: Some dependent resources of FastGPT require access to external networks, and the initialization verification process cannot be completed in an offline environment.
- Semantic retrieval fails to recall target process records, while full-text retrieval can normally recall target content. Cause: The retrieval model is not adapted to pulp and paper professional terminology, or
searchModeis only configured as semantic retrieval and does not cover matching requirements for numeric fields. - Search results include documents from other production departments that do not match the production stage of the current query. Cause:
collectionIdsis not configured correctly, and the target document collection range for retrieval is not limited.
How to Confirm the Configuration is Correct
- Upload a single pulp preparation sheet document, run a retrieval for "adjust beating degree parameters", and verify that the returned results include the corresponding field information of this document.
- Check the retrieval log to confirm that the number of returned results falls within the range specified by the
limitconfiguration, and the similarity scores fall within the interval specified by thesimilarityconfiguration. - Switch to an offline environment, test the knowledge base loading and retrieval process, and confirm that there are no stuck errors or resource loading failure prompts.
- Compare the results of semantic retrieval and hybrid retrieval, and confirm that the hybrid retrieval recall results cover both text and numeric target documents.
All configuration values listed here are common starting points, and should be validated against the user's own samples.