What this category of data looks like
Data for this category primarily comes from raw material quality certificates, heating furnace systems, press collection data, MES, non-destructive testing equipment, and customer forging drawings. Daily documents include forging process specifications, cutting lists, heating furnace records, forging process cards, mold life records, heat treatment entrustment slips, and ultrasonic testing reports. Data updates are completed for each forging, batch of materials, or furnace run. Document fields include furnace batch number, forging ratio, fiber direction, heating temperature, pressure parameters, and testing results. Some documents include quantitative testing values and process parameters.
What constraints this category’s characteristics impose on knowledge base retrieval and recall
The batch update nature of this category’s data requires retrieval to support precise filtering and recall by furnace batch number and forging batch. Documents contain a large volume of process parameters and quantitative test results, so matching must cover both structured fields and unstructured text content. Some parameters from cross-source documents use different units, so parameter normalization matching must be supported. Since errors directly lead to forging scrap and impact service safety, retrieval and recall accuracy must meet strict matching requirements to avoid missing key process or testing information.
How to set the configurations
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the accuracy requirements for forging process parameters and test results, and avoids recalling irrelevant documents with low matching scores |
limit | Top 8–12 entries | Covers all document types associated with a single batch of forgings, while controlling context length |
searchMode | mixedRecall | Balances matching needs for structured fields (such as furnace batch number) and unstructured process text |
usingReRank | Enabled | Performs secondary ranking on initial recall results to improve the accuracy of key parameter matching |
datasetSearchUsingExtensionQuery | Enabled | Adds process-related extended query terms to cover synonyms or similar process expressions |
collectionIds | Configured by document type group | Distinguishes knowledge base collections of different types such as process specifications and test reports, enabling precise targeted retrieval |
Three common mistakes
- Phenomenon: Multilingual retrieval and recall rates are extremely low. Cause: No adapted embedding model was used for multilingual scenarios, and existing knowledge base documents were not re-embedded in batches. Old embedding vectors do not match the new model.
- Phenomenon: Parameter errors are returned when calling the API to create a text collection. Cause: FastGPT v4.17.0 or a later version is not in use, or configuration parameters for the paragraph-first splitting mode were not correctly passed.
- Phenomenon: A large number of documents unrelated to forging processes appear in retrieval results. Cause: Knowledge base collections were not configured by document type group, leading to accidental recall of documents from unrelated categories.
How to confirm the configuration is properly set
- Enter the forging furnace batch number as the search term, and verify that the corresponding heating furnace records, forging process cards, and ultrasonic testing reports are included in the recall results.
- Adjust the similarity threshold, observe changes in the matching degree of recall results, and confirm that the threshold setting aligns with business requirements.
- View the splitting preview of knowledge base documents, and confirm that process parameters and steps are not split apart.
- Test multilingual retrieval scenarios, and confirm that documents with similar process expressions are correctly recalled after the extended query takes effect.
The values provided are common starting points and should be measured against the reader's own samples.