Data Profile for This Category
Data for this category comes from milk source testing, MES systems, homogenization and heat treatment equipment, online temperature and flow acquisition modules, CIP systems, and LIMS laboratory management platforms. Daily documents include raw milk acceptance records, standardized ingredient lists, pasteurization or UHT heat treatment records, CIP cleaning records, filling records, sample retention and microbiological test reports. Fields cover raw milk colony count, ingredient ratio, heat treatment temperature and duration, CIP cycle parameters, and more. Units include cfu/mL, %, ℃, seconds, MPa, and others. Data is updated in real time per milk tank and per batch, with a single batch of data linked to multi-stage production processes.
Constraints on Knowledge Base Retrieval and Recall
Real-time, high-frequency batch data updates require the retrieval chain to support low-latency recall, to avoid compliance review failures caused by data lag. Multi-source heterogeneous data formats, including structured reports, device original logs, and laboratory reports, require retrieval models to adapt to semantic matching of mixed fields, rather than relying solely on pure text keyword matching. A single batch of production is linked to multi-stage documents, requiring retrieval to support cross-document associated recall, to avoid breaking the full-link review logic from raw milk acceptance to filling. Raw materials are perishable and errors have serious consequences, requiring the relevance of retrieval results to match business risk levels, prioritizing recall of high-priority compliance verification documents.
Configuration Settings
This guide is written for FastGPT v4.17.0, and is adapted to the version’s knowledge base management and retrieval logic.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the density of professional terminology in dairy processing documents, filters low-relevance non-production documents |
limit | Top 8 entries | Covers the number of multi-stage documents linked to a single batch of production, avoiding recall of redundant data |
searchMode | mixedRecall | Combines keyword matching and semantic retrieval, adapting to mixed retrieval needs for structured reports and device logs |
usingReRank | Enabled | Performs secondary ranking on recall results, prioritizing documents strongly related to raw milk acceptance and heat treatment stages |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands industry-specific terminology for queries, such as expanding "UHT" to "ultra-high temperature instantaneous sterilization", adapting to industry terminology conventions |
collectionIds | Bind corresponding datasets by production batch | Isolate document data from different milk tanks or batches, avoiding cross-batch retrieval interference with review logic |
Common Misconfigurations
- The exported knowledge base file size is only hundreds of KB, and does not cover all production documents. The cause is that full synchronization of full-link documents such as raw milk acceptance records and microbiological test reports by production batch was not performed, and only files from some stages were uploaded.
- Imported markdown documents cannot display local images. The cause is that the associated path of local image resources was not configured, only the text content was imported, and image files were not synchronously uploaded to the knowledge base storage directory.
- Retrieval accuracy is insufficient, and heat treatment parameters required for compliance verification cannot be quickly located. The cause is that the
usingReRankconfiguration was not enabled, and thesimilaritythreshold was set too low, resulting in recall of a large number of low-relevance non-production documents.
How to Verify Configuration Correctness
- Perform cross-review retrieval for a single batch of raw milk, check that the recall results include raw milk acceptance records, heat treatment records and CIP cleaning records for the corresponding batch.
- Upload a markdown document containing local image links, confirm that the images can be loaded and displayed normally when retrieving the document.
- Trigger a full knowledge base training task, check that the training log shows the number of documents covering the full dataset, with no missing prompts.
- Export the knowledge base file, check that the file size matches the total storage capacity of all documents, with no obvious under-sizing.
The values provided are common starting points and should be measured against the reader's own samples.