What Data for This Category Looks Like
This category of data mainly comes from customer drawings, special characteristic lists, ERP/MES systems, measurement and inspection tool records, equipment process parameters, and supplier quality certificates. Daily generated documents include control plans, process flowcharts, work instructions, first-piece and patrol inspection records, SPC reports, PPAP materials, and 8D reports. The length of a single document is mostly 500–2000 characters. Data is updated per production batch, collected per shift, and synced continuously online. Most document structures have clear fields such as process numbers, special characteristic parameters, and batch identifiers. Units include commonly used industrial measurement standards such as millimeters, newton meters, and piece counts.
Constraints Imposed by These Characteristics on Knowledge Base Retrieval and Recall
The scattered data sources and high update frequency of this category require the knowledge base retrieval system to support incremental synchronization of multi-source heterogeneous data and real-time recall, to avoid affecting the timeliness of bid documents due to delayed data. Documents contain a large amount of professional content with process numbers and special characteristic parameters, so retrieval must accurately match keywords for specific process links, to prevent retrieval results from deviating from compliant documents required for bids. Subject to strict constraints from vehicle customer quality systems, recall results must be strictly associated with the corresponding customer’s special requirements, and must not mix document data from different batches or customers. At the same time, the retrieval scope must be controlled to avoid irrelevant documents interfering with quotation and document preparation.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.75–0.85 | This category of documents contains a large number of professional process parameters, so low-similarity irrelevant results need to be filtered out while retaining valid content matching special characteristics |
limit | Top 8–12 entries | Bid documents need to cover multiple types of documents. Too many single recall results will exceed context limits, while too few will fail to cover all necessary content |
searchMode | mixedRecall | Documents in this category have both semantic process descriptions and structured content such as clear numbers and parameters. Hybrid retrieval can balance accurate matching and semantic recall |
usingReRank | Enabled | Batch-recalled documents contain a large amount of similar content, and reranking can prioritize returning core documents such as PPAP and 8D reports most relevant to bid requirements |
datasetSearchUsingExtensionQuery | Enabled | Bid personnel’s queries often omit technical term prefixes. Extended queries can supplement keywords such as process numbers and special characteristics to improve recall accuracy |
collectionIds | Specify exclusive collections belonging to the bidding customer | Document data from different customers must be strictly isolated to avoid mixing customer-specific requirements, which complies with vehicle customer quality system constraints |
Three Common Misconfigurations
- Phenomenon: Retrieval results contain a large number of other batch documents unrelated to the current bidding customer. Cause: The
collectionIdsparameter is not specified, and the retrieval scope is not limited to the exclusive document collection of the corresponding customer, resulting in mixed cross-customer data. - Phenomenon: Garbled characters appear when importing online PPAP materials and SPC reports. Cause: The encoding format of the corresponding documents is not configured, and the file transfer encoding in the online environment is inconsistent with local deployment, leading to binary parsing errors.
- Phenomenon: The number of single recall results does not meet expectations, and the reference upper limit in the 100–900 range cannot be selected. Cause: The configuration rules of the
limitparameter are not correctly understood, and the reference upper limit is confused with the number of retrieval entries. The platform only provides discrete gear configuration options by default.
How to Verify Proper Configuration
- Submit a query containing a specific process number, and check whether the recall results only include documents such as control plans and work instructions from the corresponding customer.
- Import a batch of new PPAP materials, and verify whether the retrieval results can correctly match content containing special characteristic parameters.
- Adjust the
similarityparameter range, and check whether the relevance of retrieval results meets business requirements. - Check whether the
usingReRankfunction is enabled, and confirm that the sorted recall results prioritize core documents required for bids.
The values provided here are common starting points and should be evaluated against the reader’s own samples.