This configuration is based on FastGPT v4.17.0.
What data looks like for this category
Data for this category comes from ERP material master data, melting furnaces and thermometers, spectrometers, MES heat records, mold management systems, customer drawings and technical specifications. Update frequency is per heat, per batch, or daily. Daily documents include foundry process cards, batching lists, heat records, mold history logs, first article inspection records, material test reports, and nonconforming product review forms. Data fields include alloy element content values, furnace temperature, pouring parameters, defect types, mold numbers, and more. Units include degrees Celsius, seconds, kilograms, and others. Field combinations differ across various documents.
What constraints these characteristics impose on knowledge base retrieval and recall
High-frequency updated heat and batch data requires the knowledge base to support incremental synchronization, to avoid retrieving outdated production records. Multi-field combined document structures require retrieval to associate and match related fields such as alloy batching, heat, and mold number, instead of only matching single keywords. Structured numerical data requires retrieval to support numerical range queries, to meet matching needs for parameters like alloy composition and pouring temperature. Differences in fields across different documents require retrieval configurations to flexibly adjust matching rules, to cover different retrieval scenarios for process documents and exception reports.
How to set the configuration
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches high-precision numerical fields such as alloy composition and furnace temperature. Avoids introducing irrelevant results if set too low, and misses valid matches if set too high |
limit | Top 8–12 entries | A single process document or exception report needs to associate no more than 10 heat and mold data entries, to avoid overly long context |
usingReRank | Enabled | Foundry data has multi-field associations. Reranking can prioritize merging matching results from the same heat and same mold, to solve sorting anomalies |
searchMode | fullTextRecall | The query examples emphasize identifiers or specialist terms; test this mode with representative data, not as a product default. |
datasetSearchUsingExtensionQuery | Enabled | Can automatically expand numerical query ranges for batching ratios and pouring temperatures, to adapt to parameter interval descriptions in process documents |
Three common mistakes
- Enabling
usingReRankdoes not change retrieval result sorting. Some matching results from the same heat that should appear first are not prioritized. The cause is not setting fields for the same heat and same mold as reranking weight priorities, and only relying on basic similarity sorting. - When retrieving alloy batching ratios, low-match text results appear before high numerical match results. The cause is not enabling
datasetSearchUsingExtensionQuery, which cannot recognize the matching priority of numerical parameters. - A
getaddrinfo EAI_AGAIN mongoerror occurs during knowledge base incremental synchronization, and recovery fails after restarting. The cause is that MongoDB connection configuration does not adapt to the high-frequency incremental update request rate of foundries, leading to connection pool exhaustion.
How to confirm the configuration is correct
- Initiate a retrieval that includes keywords for alloy composition and furnace temperature. Verify that returned results first display associated data from the same heat. Adjust the
similarityparameter to adapt to matching accuracy as needed. - Test retrieval results after enabling
usingReRank. Confirm that matching results from the same mold and same heat are sorted as expected. - Perform an incremental synchronization operation. Check for MongoDB connection errors. Adjust system connection parameters to adapt to data update frequency as needed.
- Upload a PDF process document containing images. Confirm that the system can extract text content from images for retrieval.
The values provided here are common starting points, and should be evaluated against the reader's own samples.