What the data for this category looks like
SMT mounting and electronic assembly process data sources include BOM, Gerber, and coordinate files provided by customers, as well as real-time collected data from MES systems, mounting machines, SPI/AOI testing equipment, and reflow oven temperature collection equipment. Data update frequency follows per work order, each line change, and real-time triggers. Daily generated documents cover BOM and coordinate files, line change lists, stencil and furnace temperature records, first article confirmation reports, SPI/AOI test reports, maintenance records, and traceability labels. Fields include part number, reference designator, component batch number, coordinate offset, furnace temperature curve parameters, and more. Units include millimeters, degrees Celsius, batch numbers, and more.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Real-time and high-frequency updated process data requires the knowledge base to support incremental synchronization and fast switching, adapting to business scenarios with frequent line changes. The coexistence of multiple types of structured and semi-structured documents requires retrieval to prioritize matching exact fields such as part number and reference designator, to avoid retrieving irrelevant content. Severe errors such as wrong components and dry joints require recall results to have relatively high accuracy. A large number of work orders requires controlling the number of knowledge base entries retrieved in a single call, to avoid context overflow that affects the quality of large model responses.
How to set the configurations
These configurations are based on FastGPT v4.17.0.
Choose a retrieval mode through representative-query testing; configure category, batch, or entity filters separately because the mode itself does not enforce those boundaries.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirements for SMT process documents, and avoids retrieving irrelevant component parameters or furnace temperature data |
limit | Top 6 entries | Controls the number of knowledge base entries retrieved in a single call, adapting to the single-page context length of PCB assembly process documents |
usingReRank | Enabled | Rearranges recall results, corrects deviations from keyword matching, and improves retrieval accuracy for issues related to solder joint defect analysis |
datasetSearchUsingExtensionQuery | Enabled | Expands query terms, supplements related terms such as component substitute materials and line change steps, and covers common extended needs for process Q&A |
collectionIds | Configured according to the dataset ID bound to the work order | Associates exclusive documents such as stencil and furnace temperature records corresponding to the current work order, and avoids interference from cross-work order data |
Three common configuration mistakes
- Only the first question triggers knowledge base reference, and subsequent questions have no recall results. The cause is that the workflow is not configured with context passing, and keywords from historical conversations are not brought into the retrieval link.
- After importing a Chinese-English bilingual knowledge base, the large model does not read the corresponding Chinese content. The cause is that word segmentation optimization for Chinese corpus is not enabled, or the retrieval does not match the vector features of Chinese keywords.
- The global variable datasetid cannot be referenced in the knowledge base retrieval node. The cause is that the global variable is not mapped to the dataset ID parameter of the retrieval node, or the variable scope does not cover the current workflow node.
How to confirm correct configuration
- Enter a process-related question, and check whether the number of knowledge base entries returned by the retrieval node matches the configured
limitvalue. - View the retrieval log, and confirm that the
similarityscore of the recall results falls within the configured range. - Bind a test dataset, enter a question containing a specific part number, and verify whether the corresponding BOM or line change list content is retrieved.
- After enabling the reranking function, compare the order of recall results before and after reranking, and confirm that the result sorting aligns with business logic.
The values provided here are common starting points and should be measured against the reader's own samples.