What the data for this category looks like
Typical entities for this category are carbonated drink, juice, tea, functional beverage, and packaged drinking water factories. Daily documents include formula 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 source and raw material testing, formula systems, online sugar and acidity monitoring equipment, filling equipment, visual inspection systems, and LIMS systems. Data is updated per batch, per shift, or via continuous collection. Documents include clear fields such as feeding parameters, test values, and equipment operation records, with units including production-specific units like concentration, temperature, and filling speed.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
The data characteristics of this category impose multiple constraints on retrieval and recall. Continuously collected real-time equipment data coexists with batch-level static documents, requiring the system to support both static document retrieval and dynamic time-series data matching. Multiple document types contain specialized fields and production-specific units, requiring precise matching of parameter names and units to avoid confusion between feeding and testing indicators across different processes. Frequently updated shift and batch data require retrieval to support incremental recall mechanisms, ensuring results cover the latest records. Records from critical processes such as aseptic carrying severe risks if incorrect, requiring high-priority, highly compliant documents to be recalled first.
How to configure settings
The following configurations are based on the knowledge base retrieval and recall functionality of FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches specialized parameters and units for this category to avoid irrelevant low-similarity documents being included. Can be tightened to above 0.8 after calibration based on actual testing |
limit | Top 8–12 results | Covers key cross-process related information across multiple document types, preventing important data from being omitted in individual results |
searchMode | mixedRecall | Combines keyword and semantic retrieval to adapt to different document types such as formula sheets and equipment operation logs |
usingReRank | Enabled | Rearranges recalled results by relevance, prioritizing records from critical processes including aseptic filling and water treatment |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands synonymous phrases for user queries, adapting to colloquial questions during equipment inspection |
collectionIds | Bind dedicated datasets for corresponding production lines | Isolates documents across different production lines to avoid irrelevant cross-line data interfering with retrieval |
Three common configuration errors
- Issue: When configuring an internal Confluence dataset, page content fails to parse, and empty results or parsing failure prompts are returned. Cause: No proxy or whitelist for internal network access is configured. FastGPT only supports publicly accessible data sources by default, so internal network resources require additional access permission setup.
- Issue: When uploading doc format documents, the system indicates the format is unsupported, or content cannot be retrieved after upload. Cause: The doc format parsing plugin is not enabled, or required document conversion dependencies are not installed. Converting files to PDF format can serve as a temporary workaround.
- Issue: A large number of irrelevant water treatment records appear in retrieval results, making it impossible to accurately match queries about abnormal filling liquid levels. Cause: The
similaritythreshold is set too low, or thecollectionIdsfor the corresponding production line are not bound, leading to irrelevant cross-production-line data being recalled.
How to verify the configuration is correct
- Submit a query that includes production-specific parameters, and confirm that returned results include the fields and units of the corresponding documents.
- Switch between datasets for different production lines, and confirm that retrieval results only include content from datasets bound to the current
collectionIds. - Test access to internal network data sources, and confirm that internal network pages such as Confluence can be properly parsed and return content.
- After enabling
usingReRank, confirm that documents from critical processes are ranked at the top of retrieval results.
The values given are common starting points and should be measured against the reader's own samples.