What the data for this category looks like
Data sources for SMT mounting and electronic assembly production include BOMs, Gerber files, coordinate files provided by customers, as well as MES systems, solder paste and component batch ledgers, mounters, SPI/AOI inspection equipment, reflow oven temperature records, and more. Data is synced in real time for each work order and each line change, covering documents such as BOMs and coordinate tables, line change tables, stencil and furnace temperature records, first article confirmation reports, SPI/AOI inspection reports, maintenance records, and traceability labels. Document fields include part numbers, reference designators, component batches, furnace temperature curve parameters, coordinate values, and more. Units include millimeters, degrees Celsius, batch numbers, and others.
What constraints these characteristics impose on the knowledge base retrieval and recall link
This category of data has high real-time requirements. Retrieval must adapt to frequently updated work orders and line change data, to avoid recalling outdated stencil or furnace temperature records. Multiple types of documents coexist, including structured tables, unstructured text, and inspection reports in image format. This requires retrieval to support both keyword matching and semantic understanding. High accuracy is required for core fields: matching errors in fields such as part numbers and reference designators will directly lead to production errors, so irrelevant recall results must be restricted. In addition, the requirement to isolate data by production unit means retrieval scope must be limited to the dataset of the corresponding workshop, to avoid interference from cross-work order data.
How to set the configurations
These configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Core SMT production fields such as part numbers and reference designators require high matching accuracy to avoid irrelevant recall interfering with maintenance and inspection decisions |
limit | Top 5 | The number of documents related to a single work order typically does not exceed 5. Limiting the number of results reduces context load |
usingReRank | Enabled | Mixed data from multi-source recall requires reranking to optimize sorting results, improving matching accuracy for documents such as SPI/AOI reports and maintenance records |
datasetSearchUsingExtensionQuery | Enabled | Supplement contextual information related to work orders such as line changes and furnace temperatures, improving retrieval recall rate and adapting to multi-dimensional queries for equipment inspection |
collectionIds | Bind corresponding workshop-specific mounter, furnace temperature, and maintenance datasets | Isolate data by production unit, preventing irrelevant data from cross-workshop work orders from being recalled and ensuring retrieval accuracy |
searchMode | mixedRecall | Match both text keywords and semantics, adapting to retrieval needs for multiple types of documents such as BOM tables and maintenance records |
Three common mistakes
- Phenomenon: When initiating retrieval using a strict question-and-answer template, if the knowledge base includes stencil records in image format or first article confirmation images, "no answer found" will be returned. Cause: The strict question-and-answer template only processes plain text retrieval results by default, and does not configure parsing adaptation for image content, so it cannot extract parameter information from images.
- Phenomenon: An invalid ID prompt is returned when passing a knowledge base ID parameter during API calls. Cause: The
collectionIdsconfiguration corresponding to the SMT workshop is not bound, or the passed dataset ID is not within the authorization scope of the current workflow. - Phenomenon: When uploading a PDF maintenance record or furnace temperature report larger than 100MB, the upload progress reaches 90% and then prompts "offset out of range". Cause: The buffer parameters for large file chunked upload have not been adjusted, or network fluctuations cause chunked upload interruptions, leading to file parsing exceptions.
How to confirm the configuration is correct
- Upload a single BOM file, initiate a part number search, and verify that the returned results include component reference designators and batch information from the corresponding BOM.
- Initiate an equipment inspection query related to line changes, and check that the recalled results include stencil parameters, component replacement rules, and other content for the corresponding line change.
- Call the API interface, verify that the
similarityscore of the returned results is within the configured range, and the number of results does not exceed the configuredlimitvalue. - Upload a large-volume PDF file, confirm that the upload progress completes without an offset error, and that furnace temperature curve data within the file can be retrieved normally.
The values provided are common starting points and should be measured against the reader's own samples.