What Data for This Category Looks Like
This category’s data originates from wood pulp and waste paper batch ledgers, DCS/QCS control systems, on-machine measurement equipment, laboratory testing systems, and energy and water treatment systems. Daily documents include pulp preparation sheets, beating logs, paper machine operation records, online quality curves, laboratory reports, sheet break and grade change records. Data update frequencies cover real-time, shift-based, and per-roll production cycles. Document fields include absolute dry pulp ratio, beating degree, sheet grammage, moisture content, and some include clear physical units.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
Multi-source heterogeneous data sources require the retrieval pipeline to support cross-system associated recall, and prevent siloed production documents across departments. Real-time, shift-based, and per-roll update frequencies require recall results to cover the latest production batch data, and adapt to rapid operating condition changes. Documents include numeric fields and physical units, so retrieval configurations must support both numeric similarity matching and text semantic matching logic, to cover retrieval needs for multiple types of production documents. Event-based documents such as sheet breaks and grade changes are tightly linked to production batches, so retrieval must support filtering recall results by batch dimension, narrowing retrieval scope and improving accuracy.
Configuration Settings
These configurations are based on FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the matching accuracy requirements for pulp and paper numeric documents, and prevents low-quality invalid data from being included in retrieval results |
limit | Top 15–25 entries | Covers sufficient historical batch and current operating condition related data, to meet the context retrieval needs of production issues |
searchMode | mixedRecall | Matches both text semantics and exact keywords for numeric fields, adapting to multi-source heterogeneous production documents |
usingReRank | Enabled | Performs secondary calibration on initially recalled results, raising the priority of high-match results and adapting to complex documents such as online quality curves |
datasetSearchUsingExtensionQuery | Enabled | Expands retrieval context, associates related documents such as pulp preparation and paper machine operation records, covering the full workflow of production issues |
collectionIds | Collections specified by production department | Divides datasets by departments such as material preparation, pulping, and paper making, avoiding interference from cross-department irrelevant data in retrieval results |
Three Common Misconfigurations
- Symptom: Retrieval results have very low matching with current operating conditions, and responses deviate from core issues such as pulp preparation and paper machine operation. Cause: The
similaritythreshold was not adjusted for pulp and paper numeric documents, leading to the recall of low-match invalid data. - Symptom: Retrieval only returns Chinese content, even when an English prompt and English knowledge base documents are configured. Cause: Only Chinese documents were uploaded to the dataset, and no retrieval collection for English documents was configured synchronously.
- Symptom: Semantic retrieval scores are close to 1, but the actually recalled documents are irrelevant to the query. Cause:
usingReRankwas not enabled, and initially recalled high-score but semantically mismatched results were returned directly.
How to Verify Proper Configuration
- Upload a batch of known pulp preparation sheets and beating logs, initiate a retrieval request matching the corresponding production batch, and confirm that the recall results include the target documents.
- Adjust the
similaritythreshold, and verify that the number of recall results changes as expected relative to the threshold. - After enabling
usingReRank, compare the sorting differences between pre- and post-rerun recall results, to confirm that the reranking logic is active. - Check the field units of retrieval results, and confirm that matched documents include physical unit data required by current operating conditions.
The values provided here are common starting points, and should be evaluated against the samples appropriate for the specific use case.