What Data Looks Like for This Category
Daily documents for welded structural component manufacturing include welding procedure qualification records, welding procedure specifications, welding consumable issue records, welder qualification records, weld inspection reports, repair work orders, and more. Data sources include design drawings, material quality certificates, welding equipment parameters, welding consumable warehouse systems, non-destructive testing systems, and welder qualification ledgers. Update frequency is per weld, per batch, or per shift. Individual documents contain fields such as weld number, consumable batch number, welder qualification certificate number, and non-destructive testing results. Document formats and field dimensions vary across different sources.
What Constraints These Data Characteristics Impose on Knowledge Base Retrieval and Recall
High-frequency data updates require retrieval to cover real-time information such as latest consumable issuances and welder qualification changes, to avoid using expired data during audits. Documents with multiple types and strongly constrained fields require retrieval to support scope limitation by specific fields, for precise matching of key information like weld numbers and welder qualification certificate numbers. Audit tasks for a single batch or shift involve multiple related documents, so retrieval must recall a sufficient number of results at once, while controlling total context length to stay within processing limits. Due to the severe consequences of structural failure, retrieval results must have high relevance to reduce interference from irrelevant documents.
Configuration Settings
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Welding procedure and qualification documents require high relevance to avoid irrelevant results interfering with inspection audits |
limit | Top 8–12 results | Single batch or shift audits require coverage of multiple related documents, while controlling total context length |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
usingReRank | Enabled | Secondary calibration of relevance for multi-source heterogeneous documents, to improve recall priority of key process documents |
datasetSearchUsingExtensionQuery | Enabled | Supplement information such as weld type and equipment parameters by expanding retrieval queries, to optimize matching accuracy |
collectionIds | Bind dedicated document collections for welded structural component inspection and process work | Limit retrieval scope, exclude irrelevant documents from other categories, and improve retrieval efficiency |
Three Common Misconfigurations
- Symptom: After enabling tool calling, knowledge base retrieval returns empty results or no matching documents. Cause: Context binding configuration between tool calling and knowledge base retrieval is incorrect, and knowledge base retrieval is not included in the tool calling execution workflow.
- Symptom: Retrieval results include documents not belonging to the welded structural component category. Cause: The
collectionIdsconfiguration is not used to limit the retrieval document set, causing the retrieval scope to cover irrelevant data from other business scenarios. - Symptom: The number of documents returned by retrieval is far lower than the configured
limitvalue. Cause: Thesimilaritythreshold is set too high, filtering documents that meet business requirements, or complete documents of the target category have not been uploaded to the data source.
How to Verify Proper Configuration
- Upload specified documents such as welding procedure qualification records and welder qualification records to FastGPT, create a dedicated document collection, and record its
collectionIds. - Initiate a retrieval request that includes weld numbers and welder qualification requirements, and check whether returned results only include documents within this dedicated collection.
- Adjust the
similarityparameter, observe changes in the relevance of returned results, and confirm that the threshold meets the accuracy requirements of business audits. - Enable
usingReRankanddatasetSearchUsingExtensionQuery, initiate a test retrieval, and compare the result ranking differences before and after reranking, to confirm that the configuration takes effect.
The values provided are common starting points and should be measured against the reader's own samples.