What this category’s data includes
Data for this category primarily originates from order design systems, sheet material and wood batch ledgers, CNC cutting and drilling equipment, ERP/MES systems, and moisture content testing equipment. Daily documents include order breakdown drawings, cutting lists, edge banding and drilling programs, coating process descriptions, hardware bill of materials, trial assembly and packaging records. Both structured tabular data and drawing documents with dimension markings are present. Data updates coincide with order generation, daily production synchronization, and completion of each batch of material arrivals. Fields include panel size, material batch number, process parameters, hardware model, and others. Units involve millimeters, kilograms, cubic meters, and additional measurement units.
Constraints on knowledge base retrieval and recall
The data characteristics of this category impose multiple constraints on the retrieval and recall process. The made-to-order production model leads to strong data correlation per individual order. Retrieval must accurately match exclusive parameters such as panel size and material batch number to prevent mixing data across orders. The coexistence of multiple document types requires retrieval to adapt to both structured tabular data and unstructured process documents. Adjust chunking rules to accommodate varying data lengths across different formats. Frequently updated data sources require regular knowledge base synchronization to ensure retrieval results use the latest sheet material parameters, cutting data, and other information. The large number of panels and high parameter granularity require controlling the granularity of single retrieved data entries to avoid information overload that reduces quotation and document preparation efficiency.
Configuration Settings
All configurations below are based on FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
similarity | 0.72–0.85 | Matches precise business fields such as panel size and hardware model to cover the accuracy requirements of conventional retrieval. Adjust based on actual testing. |
limit | Top 6–8 entries | Adapts to the multiple types of segmented data that need to be cited in a single bid document, avoiding excessive results occupying context space. |
searchMode | mixedRecall | Adapts to both structured cutting lists and hardware lists, as well as unstructured order breakdown drawings and process documents, covering the multi-type document characteristics of this category. |
usingReRank | Enabled | Reorders retrieved results by relevance to improve matching accuracy for segmented parameters such as sheet material moisture content and edge banding processes. |
datasetSearchUsingExtensionQuery | Enabled | Extends queries using professional terminology for bid scenarios, optimizing retrieval effectiveness for segmented content such as order breakdown and coating. |
collectionIds | Bind corresponding datasets by business module | Separates knowledge base collections such as order breakdown data, hardware lists, and process documents to avoid cross-module data interfering with retrieval results. |
Three Common Errors
- Phenomenon: After retrieval results are returned, the large language model output includes the phrase "citation marker: [1]". Cause: The automatic splicing configuration for knowledge base citation markers is not disabled, or redundant metadata information is not filtered during retrieval.
- Phenomenon: Single retrieved data chunks are overly large, resulting in insufficient retrieval granularity. Cause: Chunking rules are not adjusted for structured data such as Excel cutting lists and hardware lists for this category. Using generic chunking parameters causes data chunks to contain too many rows of content.
- Phenomenon: Retrieval results omit specific batch sheet material parameters. Cause: Sheet material batch data in the knowledge base is not updated regularly, or the bound dataset does not include the latest ERP/MES synchronized data.
How to Verify Proper Configuration
- Upload a single set of structured cutting list and hardware list data, run a retrieval test, and confirm that the fields of the retrieved results match the source data.
- Trigger a simulated bid document preparation query, and check whether unexpected citation marker phrases appear in the large language model output.
- Manually modify a test sheet material batch data entry, synchronize the knowledge base update, run a retrieval, and confirm that the retrieved results include the latest modified content.
- Adjust the
similarityparameter range, compare the number of retrieval results and relevance across different values, and confirm that the matching accuracy meets business requirements.
All provided configuration values are common starting points, and should be evaluated against the reader's own sample data.