Data Profile for This Category
Production entities include carbonated beverage, juice, tea, functional beverage, and packaged drinking water factories. Daily production documents include recipe sheets, water treatment records, blending records, sterilization or aseptic filling records, packaging material inspection reports, and finished product inspection reports. Data is sourced from water and raw material testing, recipe systems, online sugar and acidity monitoring equipment, filling equipment, visual inspection systems, and LIMS systems. Update cadence follows: adjust recipe parameters for each production batch, generate complete water treatment and filling records per shift, and continuously collect and update online detection data. Document fields include concentrate dosage, sterilization temperature, filling liquid level threshold, packaging material batch number, and similar fields. No fictional statistical data or percentage statements are used.
Constraints for Knowledge Base Retrieval and Recall
Multiple heterogeneous data sources require the retrieval process to support cross-system associated recall. Relying solely on static documents will miss real-time online monitoring parameters. The high-frequency update cadence requires the knowledge base to have an incremental update mechanism. This ensures production records generated per batch and shift, and continuously collected online data, are synchronized in a timely manner. This prevents recalling outdated process parameters. Most documents focus on production parameters and inspection results. Retrieval must accurately match fields and units to avoid parameter deviations caused by fuzzy matching. Frequent product switching scenarios require retrieval to split datasets by category. This ensures recall results only correspond to the currently produced beverage category.
Configuration Settings
This configuration is based on the knowledge base retrieval module of FastGPT v4.17.0.
Choose a retrieval mode through representative-query testing; configure category, batch, or entity filters separately because the mode itself does not enforce those boundaries.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precision requirement for matching production parameter text, and avoids recalling irrelevant cross-category process documents |
limit | Top 8 entries | Covers core process links for a single production batch, while controlling the total context length |
usingReRank | Enabled, re-rank top 10 results | Performs secondary semantic ranking on initially recalled results, and improves the display priority of highly relevant documents |
datasetSearchUsingExtensionQuery | Enabled, extend query length to 100–150 characters | Covers associated parameters in production documents, and supplements contextual information of the process chain |
collectionIds | Collection IDs split by production category | Isolates process documents of different beverage categories, and avoids interference from cross-category retrieval |
Three Common Misconfigurations
- Phenomenon: Retrieval results are directly output on the interface, and subsequent AI answers repeatedly display the results. Cause: The direct rendering switch for retrieval results is not turned off. This causes the retrieval and answer stages to repeatedly output recalled content.
- Phenomenon: Retrieval fails to recall the top-ranked knowledge base document. Cause: The
usingReRankre-ranking function is not enabled. The initial retrieval similarity ranking does not perform secondary calibration based on text semantics, leading to highly relevant documents being ranked lower. - Phenomenon: Knowledge base file retrieval performance is poor, and cannot accurately match process parameters. Cause: Documents are not split by process links, and multiple types of content such as recipes, water treatment, and filling are mixed. This interrupts semantic association after chunking.
How to Verify Proper Configuration
- Submit a query that includes core process parameters, and check whether the source of retrieval results belongs to the currently configured dataset, with no cross-category documents included.
- Adjust the value range of
similarity, and verify whether the number and relevance of retrieval results meet the needs of the production scenario. - After enabling the
usingReRankfunction, compare the ranking of recall results before and after re-ranking, and confirm whether the position of highly relevant documents has improved. - Submit batch queries, check whether there are duplicate retrieval results, and confirm that the rendering switch has been correctly configured.
The values provided are common starting points and should be evaluated against the reader's own samples.