What this category of data looks like
Data for this category comes from procurement orders, drawing specifications, supplier quality certificates, inspection equipment, historical quality data, and supplier portals. Daily documents include inspection specifications, sampling plans, incoming material reports, supplier audit reports, exception contact sheets, 8D reports, and concession acceptance approval forms. Documents typically include fields such as batch number, supplier code, inspection items, judgment results, sampling quantity, and number of non-conforming items. The update rhythm is once per incoming batch, supplier evaluation is conducted monthly or quarterly, and audits are completed irregularly.
What constraints do these characteristics impose on the knowledge base retrieval and recall link
Data for this category is updated frequently by batch and supplier. Retrieval must accurately match fields such as batch number and supplier code to avoid mistakenly recalling outdated specifications across batches. There are diverse daily document types, with significant differences across scenarios such as inspection specifications and exception contact sheets. Retrieval needs to filter recall results by document type to ensure retrieved content matches the current bid scenario. Data fields include quality indicators such as inspection items, sampling plans, and non-conforming judgments, so retrieval must cover these specific fields. Field indexes need to be configured to accurately locate historical quality data for corresponding suppliers. Additionally, data sources are scattered, so a unified index structure is required to quickly locate quality documents for the corresponding batch during retrieval.
How to set the configuration
This configuration is based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Matches the density of professional terminology in this category, filters low-relevance quality text |
limit | Top 8 entries | Covers the number of recall results per retrieval, adapts to the multi-document reference needs of bid documents |
searchMode | embedding | Focuses on structured quality data retrieval for this category, avoids interference from irrelevant content |
usingReRank | Enabled | Rearranges recall results by business relevance, prioritizes core documents such as inspection specifications and sampling plans |
datasetSearchUsingExtensionQuery | Enabled | Expands the retrieval scope to cover associated queries for fields like supplier codes and batch numbers |
collectionIds | Use the supplier quality document collection ID for the current project | Limits the retrieval scope to the exclusive document collection for this category, prevents mixing data from other categories |
Three common mistakes
- Retrieval returns 0 results, caused by incorrect
collectionIdsconfiguration, which limits the wrong document collection and does not include the supplier quality documents required for the current bid. - Imported Chinese inspection specifications display garbled text, caused by the document encoding not being set to UTF-8 and not being converted to a text format compatible with FastGPT.
- Knowledge base disk usage exceeds expectations, caused by not configuring automatic cleanup of outdated data such as incoming material reports and 8D reports, and not deleting redundant split blocks and embedding vectors.
How to confirm the configuration is correct
- Upload a latest supplier audit report, enter the corresponding inspection item in the retrieval interface, and check whether the corresponding inspection specification document can be matched.
- Import a Chinese-format sampling plan document, check that the interface displays no garbled text, confirming that the encoding settings are correct.
- View the storage statistics of the knowledge base, confirm that only three types of data are included: original files, split file blocks, and embedding vectors.
- Adjust the
similaritythreshold, test the number of recall results under different thresholds, and confirm that the matching degree meets business requirements.
The values provided are common starting points and should be measured against the reader's own samples.