What the data for this category looks like
Data for this category originates from engineering orders, sand, gravel, cement and admixture testing, aggregate moisture content testing, mixing plant control systems, and project laboratories. Daily documents include mix proportion notices, opening identifications, production and delivery slips, raw material inspection records, test block ledgers, and strength reports. Updates occur per vehicle production batch, per raw material incoming batch, and daily summary statistics. Document fields include design strength, aggregate dosage, departure time, pouring location, and similar items. No unsupported statistical values or percentage statements are included.
Constraints imposed on vector models and indexing workflows
Indexes must support incremental synchronization for frequently updated production and raw material data, eliminating performance loss from full reconstruction. Indexes must support vectorization adaptation for both structured numeric fields and unstructured text content to handle mixed document types. Indexes must support connection with documents in different formats exported from multiple systems to accommodate decentralized data sources, reducing preprocessing costs. Indexes must support group matching based on fields such as batch number and vehicle number to meet associated retrieval requirements for single data entries, improving retrieval accuracy. Indexes must retain field-level association relationships to preserve the strong link between test block strength and age, avoiding split business logic during retrieval. The defaultToken setting must match the maximum length limit of a single text block to adapt to vectorization processing of ready-mixed concrete documents.
How to configure the settings
This configuration is adapted for FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the average length of documents such as mix proportion notices and test block records, avoiding excessive splitting or redundancy of single chunk content |
| Overlap Length | 50–100 characters | Retains business association information across chunks, ensuring mix proportion parameters do not lose context when spanning chunks |
| Similarity Threshold | Calibrated via actual testing | Adapts to semantic matching accuracy for ready-mixed concrete production data, avoiding false recall of unrelated raw material inspection records |
defaultToken | 2048 | Matches the maximum token limit for a single text chunk, adapting to input requirements of Chinese vectorization models |
maxToken | Set to the selected embedding model/service maximum supported input tokens | maxToken is the input-token maximum supported by the selected embedding model/service; verify it in that provider configuration. See the v4.17.2 model-field reference. It is not retrieval/LLM context or a FastGPT-wide fixed cap. |
| Single Search Limit | Top 10 entries | Controls the number of retrieval return results, avoiding redundant data interfering with the auditing process |
Three common configuration mistakes
- Phenomenon: Semantic matching scores of retrieval results remain consistently high, with no obvious improvement after switching vector models. Cause: Differentiated weights are not configured for structured numeric fields such as slump and strength grade for ready-mixed concrete, causing semantic matching of unrelated texts to disrupt overall scores.
- Phenomenon: Auxiliary data cannot be retrieved and matched. Cause: Vectorization configuration for auxiliary data is not enabled, or auxiliary data is not bound to the index group of the corresponding inspection and auditing task.
- Phenomenon: Multiple vector groups generated from a single production record cannot be recalled via batch association. Cause: Vector grouping rules based on business fields such as vehicle number and batch number are not configured, leading to scattered storage of multiple vector groups and inability to accurately match all associated data of the corresponding batch.
How to verify correct configuration
- Upload a single mix proportion notice, retrieve and match the production and delivery slip of the same batch, and verify the business relevance of the returned results.
- Trigger an incremental indexing task, check whether newly uploaded raw material inspection records complete vectorization and index update within a reasonable time frame.
- Adjust the similarity threshold parameter, observe changes in the number of retrieval recall results, and confirm that the threshold meets the accuracy requirements of business auditing.
- Verify the multi-vector grouping configuration, retrieve production data for a specified vehicle number, and confirm that the returned results include all associated documents of that batch.
The values provided are common starting points and should be measured against the reader's own samples.