What the data for this category looks like
Data sources for plastic injection molding scenarios include resin batch property sheets, injection molding machine operating parameters, mold temperature controller data, MES system records, and dimensional and appearance inspection reports. Daily documents cover injection molding process cards, raw material drying records, mold history logs, first-piece inspection forms, patrol inspection records, and material change and machine cleaning records. Data update frequency follows per production batch, each mold or material change, and each shift. Document fields include mold number, resin model, drying temperature and duration, filling, holding pressure and cooling parameters, and detected dimensional tolerances. Units include degrees Celsius, seconds, megapascals, millimeters, and others.
What constraints these characteristics impose on the knowledge base retrieval and recall link
High-frequency data updates require the retrieval link to support incremental synchronization. It adapts to the document update rhythm of each mold or material change. Multi-field document structures with clear process parameters require retrieval to balance semantic matching and precise keyword matching. This avoids recalling unrelated non-process data. Business characteristics with severe error consequences require recalled data to have high confidence. Low-relevance results must be filtered out. Data from multiple systems requires retrieval to be limited to the exclusive document collection for plastic injection molding. This avoids mixing in document data from other categories.
How to set the configurations
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the precision requirement for matching injection molding process parameters, avoiding recall of unrelated mold maintenance or raw material procurement documents |
limit | Top 6–8 results | The volume of process-related documents per batch is moderate. Limiting the number of recalled results avoids context overflow while covering key process scenarios |
searchMode | mixedRecall | Balances semantic matching for process parameters and precise keyword matching for mold number, resin model and other fields, adapting to multi-field retrieval needs |
usingReRank | Enabled | Minor differences in injection molding process parameters directly affect result validity. Re-ranking filters low-relevance recalled results |
datasetSearchUsingExtensionQuery | Enabled | Expands user process questions (such as "adjustment methods for shrinkage defects") into associated queries for drying parameters and mold temperature parameters, matching more relevant documents |
collectionIds | Specify collection IDs for injection molding processes, mold history logs and raw material records | Limits the retrieval scope, avoiding mixing in document data from other categories and ensuring recalled results only target plastic injection molding scenarios |
Three common mistakes
- Phenomenon: The knowledge base can retrieve matching documents, but the large model does not organize and output results. Cause:
datasetSearchUsingExtensionQueryis not enabled, resulting in recalled data not being correctly associated with the context of the user's question, or thelimitvalue is set too low, resulting in insufficient recalled data to support a complete answer. - Phenomenon: Single-document retrieval takes more than 15 seconds. Cause:
collectionIdsis not specified, causing the retrieval scope to cover all platform documents, or thesimilaritythreshold is set too low, recalling a large number of irrelevant data and increasing processing time. - Phenomenon: After deploying FastGPT v4.17.0 offline, the knowledge base query function malfunctions. Cause: Local embedding model dependencies are not configured, or offline injection molding process document collections are not properly synchronized, resulting in failure to load locally stored exclusive data during retrieval.
How to confirm the configuration is correct
- Upload a single injection molding process card, initiate a query for the corresponding parameters, and verify that the recalled results include the core fields of the document (such as filling pressure, cooling time).
- Adjust the
similaritythreshold, verify that the relevance of recalled results changes as expected with the threshold, and the precision of recalled results improves as the threshold increases. - After enabling
usingReRank, compare the number of recalled results before and after enabling to confirm that the re-ranking logic takes effect and filters low-relevance results. - Disconnect the external network in an offline environment, initiate a process question query, and confirm that locally stored injection molding document data can still be recalled normally.
Note: The configuration values listed above are common starting points. Validate them against your own specific samples and use cases.