This document is written based on FastGPT v4.17.0.
What the Data for This Category Looks Like
The data sources for PCB manufacturing scenarios include Gerber files, drill files, BOM lists delivered by customers, as well as engineering production instructions, drilling/lamination parameters, bath solution test records, AOI inspection reports, electrical testing reports and slice analysis reports generated by the factory’s internal CAM. Data is updated by production batch and work team. Some equipment parameters and testing data are collected continuously. The document structure includes fields such as process parameters, testing results, batch association identifiers, and some are proprietary format engineering files without a unified standardized text format.
What Constraints Do These Characteristics Impose on the Knowledge Base Retrieval and Recall Link
Multi-source and heterogeneous data sources (customer original files, internal inspection reports, equipment parameters) require retrieval to adapt to multiple formats, avoiding recall bias caused by only supporting plain text. High-frequency updates by production batch and work team require the knowledge base incremental update mechanism to adapt to real-time needs, preventing recall of outdated data. The non-standardized structure of proprietary engineering files requires retrieval to retain original field associations, avoiding excessive splitting that leads to loss of process parameters. The large volume of continuously collected equipment data requires limiting the context length of single-batch retrieval to avoid exceeding model processing limits. At the same time, engineering data conversion and electrical connectivity are core links. Retrieval must accurately associate historical records of key processes such as drilling and lamination to ensure audit accuracy.
How to Configure the Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.80 | Matches the precision requirement for PCB process parameters, avoiding recall of unrelated lamination or drilling parameters |
limit | Top 6–8 entries | Adapts to the volume of process documents for a single production batch, avoiding context overload that affects model inference |
searchMode | fullTextRecall | Use full-text retrieval for text content. Field, metadata, batch, and document-type constraints require separate filters and are not enforced by this mode. |
usingReRank | Enabled | Performs secondary ranking on multi-source heterogeneous inspection reports and parameter documents, improving the recall priority of abnormal process records |
datasetSearchUsingExtensionQuery | Enabled | Expands associated queries for customer original materials, covering potential parameter needs in the engineering conversion link |
collectionIds | Knowledge base collection bound by production batch | Isolates process data from different customers or batches, avoiding parameter confusion across batches |
Three Common Misconfigurations
- Calling the knowledge base retrieval interface returns an unsupported format error, but the corresponding file link can be downloaded normally via browser for docx format documents. Cause: No parsing adaptation is configured for proprietary format engineering files, and only plain text format recall processing is supported.
- The number of retrieval results does not match the configured
limitparameter, with too many or too few recall results. Cause: ThecollectionIdsparameter is not bound, and documents across batches are recalled simultaneously, leading to results exceeding expectations or missing target batch data. - Retrieval results do not associate key process parameters for drilling or lamination, with empty fields or irrelevant content. Cause:
datasetSearchUsingExtensionQueryis not enabled, and associated queries for customer original materials are not expanded, leading to failure to match hidden parameter needs in the engineering conversion link.
How to Confirm the Configuration Is Correct
- Upload the Gerber analysis records and AOI reports for a single production batch, perform retrieval, and check whether the returned results include the exclusive process parameters of this batch.
- Adjust the
similarityparameter range, verify whether the matching precision of retrieval results meets the needs of process auditing. - After enabling
usingReRank, check whether the sorted results place abnormal inspection reports (such as hole copper abnormalities) in prominent positions. - Check whether
collectionIdsis bound to the knowledge base collection of the current production batch, confirm that documents across batches are not recalled.
The values provided here are common starting points and should be evaluated against the reader's own samples.