What this category of data looks like
This category of data mainly comes from customer product models, mold flow/forming simulation files, CAD/CAM source files, machining and measurement records, and sample test part data. Daily documents include mold design task sheets, DFM reports, part drawings, processing technology cards, mold trial reports, mold modification notices, and mold history records. Data updates are carried out on a per-project basis, and are synchronized after each mold trial or modification. Most document structures include fields such as mold number, project association identifier, processing parameters, and defect records. Some drawing files are in binary format.
What constraints these characteristics impose on the knowledge base retrieval and recall link
Multi-source and heterogeneous data sources require the retrieval system to support multiple formats such as text, CAD drawings, and simulation results, and adapt to the vector modeling logic of different data types. Following the update rhythm triggered by projects, mold trials, and modifications, the retrieval link needs to support fast incremental synchronization to avoid retrieving outdated data. Documents contain a large number of professional engineering fields, and the field association relationship must be retained during retrieval to avoid losing context after splitting. The long-term cumulative nature of mold history records requires retrieval results to be linked to historical mold modification records to assist in current problem troubleshooting.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Mold design documents have high professionalism, requiring filtering of low-relevance results. This range balances retrieval completeness and accuracy, and is compatible with the vector retrieval logic of FastGPT v4.17.0 |
limit | Top 6 entries | Production ledger summary needs to balance information comprehensiveness and context length. 6 results can cover core processing and mold modification records |
usingReRank | Enabled | Short texts such as mold trial reports and mold modification notices are prone to similarity ambiguity, and re-ranking can improve the accuracy of result sorting |
datasetSearchUsingExtensionQuery | Enabled | Engineering terms often have abbreviations and standard name correspondences, and extended queries can cover more relevant documents |
searchMode | mixedRecall | Combines full-text and vector retrieval to adapt to different retrieval needs of drawing text and structured parameters |
Three common mistakes
- Phenomenon: Retrieval results return unrelated mold trial reports, and the mold number field is missing. Cause:
datasetSearchUsingExtensionQueryis not enabled, and retrieval is not performed by associating project identifier fields. - Phenomenon: An error indicating vector model loading failure is prompted when calling the knowledge base. Cause: The correct vector model configuration is not specified, or the model file path configuration is incorrect.
- Phenomenon: After calling MySQL ledger data via Function CALL, the large model cannot reference the original text fragments. Cause: The database query results are not passed as knowledge base context, and the retrieval result original text display configuration is not enabled.
How to confirm the configuration is correct
- Upload a latest mold modification notice document, perform retrieval, and check whether the returned results include this document and the associated mold trial report.
- Adjust the
similarityparameter range, observe the change in the relevance of retrieval results, and confirm that the impact of the parameter on the results meets expectations. - After enabling
usingReRank, compare the result sorting before and after re-ranking to confirm that the re-ranking function works properly. - Verify that the MySQL data returned by Function CALL can be directly referenced by the large model, and confirm that the context delivery link is complete.
The values provided here are common starting points, and should be evaluated against the reader's own samples.