What this category of data looks like
The business data for PCB manufacturing mainly comes from Gerber, drill list, and BOM files delivered by customers, plus CAM output materials, production equipment parameters, bath solution test results, AOI inspection reports, electrical test reports, and laboratory slice reports. Data is updated per production batch, generated at each shift frequency, or continuously collected by equipment. Daily documents include engineering production instructions, drilling lamination parameter records, chemical analysis ledgers, AOI and electrical test reports, and more. Most fields contain process parameters, test values, and batch identifiers. There is no unified general format, so documents must be stored classified by production stage.
What constraints these characteristics impose on the knowledge base retrieval and recall link
PCB manufacturing data is multi-source and heterogeneous, including structured process parameters and unstructured inspection reports. This requires retrieval to support both structured field matching and full-text semantic recall. High-frequency batch updates and continuously collected device data create a need for incremental index synchronization, to avoid excessive computing resource usage from full index rebuilding. Specialized process fields and non-uniform document formats require retrieval to support filtering by production stage and batch identifiers, to narrow the recall scope. The timeliness requirement for abnormal report drafting requires controlling retrieval response time to match the waiting window of the production stage.
How to Configure the Settings
Reference configuration options under FastGPT v4.17.0:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the precision requirement for PCB process parameters, filters out low-correlation non-process documents |
limit | Top 8 entries | Controls the number of documents recalled per request, avoids context overflow, and adapts to the compact nature of process documents |
searchMode | mixedRecall | Combines keyword and semantic retrieval, adapts to the mixed retrieval needs of structured parameters and unstructured reports |
usingReRank | Enabled | Rearranges recall results to prioritize returning abnormal reports or parameter documents that best match the current process issue |
datasetSearchUsingExtensionQuery | Enabled | Automatically expands retrieval terms, covers synonyms and abbreviations of specialized terms such as Gerber numbers and drilling parameters, and improves recall accuracy in professional scenarios |
collectionIds | Grouped by production stage | Limits the retrieval scope to the document set of the current process step, narrows the retrieval scope and improves speed |
Three Common Configuration Mistakes
- Phenomenon: Knowledge base retrieval time exceeds the preset waiting window, failing to meet production stage rhythm requirements. Cause: The
collectionIdsconfiguration was not used to limit the retrieval scope by production stage, resulting in full-library retrieval that matches a large number of irrelevant documents. - Phenomenon: GPT-4o-mini call error logs appear during workflow operation. Cause: The
limitconfiguration was not set reasonably, leading to too many recalled documents, exceeding the model's token limit for context, and triggering additional model calls. - Phenomenon: The uploaded process document directory structure is automatically deleted, and duplicate content is cleaned up. Cause: The automatic document cleaning function of the knowledge base was not turned off. The default rules remove hierarchical directories and duplicate text blocks.
How to Verify the Configuration is Correct
- Submit a retrieval request containing specific process parameters, and check whether the
similarityscores of the recalled results fall within the preset range. - View retrieval time records to confirm they meet the waiting requirements of the current production stage. Adjust the
limitvalue or the on/off state ofusingReRankto optimize performance. - Upload a test document with a directory structure, and confirm that the directory structure is not automatically modified.
- Trigger incremental data synchronization, confirm that newly uploaded batch documents can be recalled quickly, and there is no prompt for full index rebuilding.
The values provided are common starting points and should be measured against the reader's own samples.