What the data for this category looks like
Mold design and manufacturing daily documents include mold design task sheets, DFM reports, part drawings, machining process cards, mold trial reports, mold modification notices, and mold records. Data sources include customer product models, mold flow or forming simulation results, CAD/CAM files, machining and measurement data, and mold trial sample information. Data is updated as projects progress, with synchronized updates after each mold trial or modification. Documents contain fields such as process parameters, component dimensions, material identifiers, and defect records, with units including millimeters, material grades, machining man-hours, and more.
What constraints these characteristics impose on knowledge base retrieval and recall
The mixed multi-type documents for mold design and manufacturing include both structured process parameters and unstructured mold trial descriptions. Retrieval and recall must balance semantic matching and keyword matching. Data is updated based on projects and mold trials or modifications, so the latest versions of document fragments must be prioritized for recall, and outdated mold modification notices or mold trial reports must be excluded. Bidding scenarios require association with project-specific mold types, so recall results must be filtered through project metadata to prevent mold data from unrelated projects from being included. Additionally, documents contain precise dimensions and process parameters, so the precision of recall results must be controlled to avoid irrelevant low-match fragments interfering with quotation and document preparation.
How to set the configurations
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Mold data includes precise process parameters; this range balances recall coverage and precision, filtering out irrelevant low-match fragments |
limit | Top 8 entries | Bid documents need to cover multi-dimensional process and mold trial data; 8 entries provide sufficient context without exceeding conventional context window limits |
searchMode | mixedRecall | Balances semantic matching and keyword matching, adapting to mold documents that contain both process descriptions and clear parameter numbers |
usingReRank | Enabled | Among multiple recall results, reorder by relevance to the current bidding project, preventing low-relevance old mold trial reports from being returned first |
datasetSearchUsingExtensionQuery | Enabled | Bidding questions often include project-specific mold type descriptions; query expansion can cover more relevant document fragments |
collectionIds | Knowledge base collection IDs bound to the project | Isolate mold data for different bidding projects, preventing recall of documents from unrelated projects |
Three Common Errors
- Recall results do not include the corresponding document file names, making it impossible to quickly locate the original reference materials. This occurs because the retrieval parameters for returning document metadata are not configured, and the recall switch associated with file names is not enabled.
- The response returned by the API call does not include knowledge base matching content. This occurs because the corresponding mold project's knowledge base collection is not bound in the chat application, or the correct
collectionIdsare not specified in the request parameters. - The retrieval test interface cannot directly edit index parameters. In FastGPT v4.17.0, this edit button has been moved to the knowledge base content management page; index debugging must be completed by entering the corresponding document list.
How to Confirm the Configuration Is Set Properly
- Launch a retrieval test, and check whether the returned results include document file names and affiliated project information.
- Pass the specified
collectionIdswhen calling the API interface, and verify that the returned results include corresponding document fragments such as mold process cards and mold trial reports. - Adjust the
similaritythreshold, observe changes in the precision of recall results, and confirm that the threshold meets business requirements. - After enabling
usingReRank, verify whether the returned results are sorted to prioritize the latest mold trial data for the current bidding project.
The values provided here are common starting points and should be evaluated against the reader's own samples.