What Data for This Category Looks Like
Data for this category comes from ERP material master data, melting temperature measurement equipment, spectral analyzers, MES systems, mold management systems, and customer technical drawings. Daily documents include foundry process cards, batching lists, heat records, mold histories, first-piece inspection records, material reports, and nonconforming product review forms. Data updates are completed per heat, per batch, or daily. Documents include fields such as heat number, pouring temperature, alloy batching parameters, and mold usage count. These documents do not have a unified fixed format, and are dynamically generated along production links.
How These Characteristics Impact Knowledge Base Retrieval and Recall
Multi-source heterogeneous data sources require unified field mapping rules during retrieval to avoid matching confusion across system data. The high-frequency update rhythm requires the recall logic to adapt to dynamically generated heat and batch data, preventing expired process or quotation information from being called. There are many professional fields with close business connections, so the recall scope must be strictly limited to exclusive documents for the foundry category to avoid irrelevant content interfering with results. Bid scenarios rely on precise matching of historical processes and quotations, so the relevance requirement for recall is higher than general retrieval, and non-target documents must be filtered.
How to Set Configurations
This page applies to FastGPT v4.17.0. The following configuration items are set based on the functional logic of this version.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
similarity | 0.72–0.82 | Meets the precision requirements for matching foundry professional terminology and business parameters, balancing recall coverage and relevance |
limit | Top 10–15 entries | Controls the number of documents recalled per single request, avoiding exceeding the context carrying limit of bid documents |
usingReRank | Enabled | Rearranges recall results by business relevance, prioritizing display of core documents related to heats and batching |
datasetSearchUsingExtensionQuery | Enabled | Expands professional term queries to cover specialized expressions in the foundry field, such as associating "shrinkage porosity" with internal defect review forms |
collectionIds | Only include knowledge base collection IDs for foundry processes, quotations, and heat records | Limits the recall scope to exclude documents from other categories, preventing irrelevant content from being included |
searchMode | embedding | Adapts to the semantic characteristics of foundry professional documents, matching the association logic between process parameters and historical quotations |
Three Common Mistakes
- Phenomenon: Non-foundry category documents are mixed into retrieval results, or heat records irrelevant to the current bid project are recalled. Cause: The scope of
collectionIdsis not limited, resulting in content from other knowledge base collections being recalled. - Phenomenon: Garbled characters appear in the knowledge base after online deployment, with no abnormalities in local deployment. Cause: The online environment’s file encoding does not adapt to the professional symbols in foundry documents, and the encoding format is not uniformly set.
- Phenomenon: It is impossible to select a
limitparameter value between 100 and 900, only 100 or 900 gear options are available. Cause: The current FastGPT version’s interface configuration item does not allow custom input for this interval, only fixed gear options are provided.
How to Confirm Configurations Are Properly Set
- Input a foundry professional question, such as "Alloy batching parameters and pouring temperature for a certain model of casting", and check whether the recall results only include documents from the foundry category.
- Check the number of retrieval results to confirm they match the range set by the
limitparameter. - Check whether garbled characters or extra separators appear in the response, confirming that format filtering has taken effect.
- Compare results from different retrieval requests to confirm that the rearrangement logic prioritizes display of core business documents.
The values provided here are common starting points and should be evaluated against the reader's own samples.