What data for this category looks like
Data for this category typically comes from order design systems, sheet and wood batch ledgers, CNC cutting and drilling equipment, ERP/MES systems, and moisture content testing equipment. Daily documents include cut-to-size drawings, cutting lists, edge banding and drilling programs, coating process specifications, hardware bill of materials, trial assembly and packaging records. Data updates are completed alongside order generation, daily production progress updates, or each batch of material arrival. Fields include panel dimensions, sheet batch numbers, moisture content values, processing times, hardware specifications, and more. Corresponding units are millimeters, batch numbers, percentages, timestamps, and model identifiers.
What constraints do these characteristics impose on knowledge base retrieval and recall
The custom order-driven nature of this category results in a large number of panels and highly personalized parameters. Retrieval must accurately match unique identifiers such as panel numbers and sheet batch numbers to prevent mismatches caused by generalized matching. The coexistence of multiple document types requires the retrieval link to support parsing of content in different formats, including drawings, program text, and list tables. The high-frequency update rhythm requires the knowledge base synchronization mechanism to match the update frequency of each order, daily, or each batch. This prevents data lag from affecting the accuracy of ledger summary. Core parameters such as panel dimensions and moisture content directly impact production results. Retrieval must prioritize recalling content strongly related to the current production task to reduce interference from irrelevant results.
How to configure settings
The following configurations are adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
similarity | 0.75–0.85 | Meets the need to accurately match fields such as panel parameters and batch numbers for this category. Avoids irrelevant recall caused by too low a value, and missing valid content caused by too high a value |
limit | Top 8 entries | This category has a large number of panels per single order. A single recall sufficiently covers core production parameters while controlling context length |
searchMode | mixedRecall | Combines semantic/vector and full-text retrieval for diagrams and process text; exact size or batch filters must be configured separately |
usingReRank | Enabled | For scenarios with many customized parameters, reranking optimizes the ranking of highly relevant results and reduces misjudgments |
datasetSearchUsingExtensionQuery | Enabled | Cooperates with the question optimization step in the workflow to expand retrieval keywords and cover detailed fields such as panel numbers and hardware specifications |
collectionIds | Bind the knowledge base collection associated with the current order | Accurately limits the retrieval scope, avoids interference from cross-order data, and adapts to the independent data requirements of customized orders |
Three common configuration mistakes
- After importing documents such as cutting lists and cut-to-size drawings, the dataset status remains showing as indexing for a long time and does not enter the ready state. Cause: No parsing trigger rule is configured for multi-format documents of this category, or the knowledge base synchronization frequency does not match the update rhythm of each order or each batch.
- After switching the vector model, the retrieval returns a similarity value of 10000+, and valid filtering thresholds cannot be set. Cause: The similarity value range of this model is not clarified, and the general 0-1 range filtering rule is directly applied without adjusting to the valid range of the corresponding model.
- The workflow ends early after reaching the knowledge base search link and cannot enter the AI conversation step. Cause: The recalled content exceeds the context length limit, or
datasetSearchUsingExtensionQueryis not enabled, leading to keyword matching failure and abnormal termination of the workflow.
How to confirm the configuration is correct
- Import a batch of cutting lists and cut-to-size drawings corresponding to an order, wait for synchronization to complete, and check whether the knowledge base parsing status shows ready.
- Enter a test query containing panel dimensions and sheet batch number, verify the ranking and relevance of retrieval results, and adjust the
similarityandusingReRankconfigurations to achieve the expected recall effect. - Trigger a workflow execution test, confirm that the entire process from question optimization, knowledge base search to AI conversation completes normally without early termination.
- Check the number of results returned by retrieval, confirm that it matches the
limitconfiguration requirements, with neither too many nor too few recalled contents.
The values provided here are common starting points and should be measured against the reader's own samples.