What Data Looks Like for This Category
The primary data sources for this category are ready-mixed concrete batching plants, project laboratories, and quality scheduling departments. Daily documentation includes mix design notifications, batch opening certifications, production and delivery slips, raw material inspection records, test block test records, strength reports, and more. Data sources cover engineering orders, sand, gravel, cement, and admixture inspections, aggregate moisture content tests, batching plant control systems, and laboratory testing systems. Updates roll out for individual batch raw materials, each vehicle’s production task, and daily test results. Document fields include batch numbers, test values, timestamps, operation identifiers, and more. Units align with document type: kilograms for quality-related data, megapascals for strength-related data, with matching units for other document types.
What Constraints Do These Characteristics Impose on the Vector Models and Indexing Step
High-frequency updates require indexes to support incremental generation, avoiding full reconstruction that consumes excessive resources. Mixed input from multiple document types requires vector models to support both structured test data and unstructured text content. Wide variance in document lengths — from short mix design notifications to multi-page strength reports — demands flexible chunking strategies. Review accuracy for critical parameters such as aggregate moisture content and slump directly impacts production quality, requiring indexed retrieved documents to have high relevance and avoid missing core logic. Differences in data formats across multiple sources require indexes to adapt to document structures exported from different systems.
Configuration Settings
These configurations are based on FastGPT v4.17.0, and involve two core parameters: defaultToken and maxToken. The specific configuration items and recommendations are as follows:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to documents of varying lengths such as mix design notifications and test block strength reports, and avoids breaking the associated logic of critical parameters |
| Overlap Length | 50–100 characters | Connects associated information across adjacent chunks, such as production logic before and after aggregate moisture content corrections |
| Similarity Threshold | Calibrated via actual testing | Matches the accuracy review requirements for critical indicators such as slump and strength in ready-mixed concrete production |
| Single Search Limit | Top 10 results | Controls the number of returned results, avoiding redundant information that interferes with judgments during technical cross-review |
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. |
defaultToken | 500 | Adapts to the base token consumption of single-document chunking, and avoids exceeding model input limits |
Three Common Mistakes
- Issue: An error indicating model mismatch is returned when calling the knowledge base index generation interface, with error text containing
undefined model must match "^(text. Cause: No embedding model supported by the platform is selected in the configuration, or an undefined model name is entered. - Issue: After replacing the embedding model, the previously imported knowledge base cannot reuse existing indexes, and vector data must be regenerated. Cause: Different embedding models generate different vector spaces, and existing indexes cannot match the vector features of the new model.
- Issue: The sorting of knowledge base search results does not match expected similarity. Cause: Reranking configuration is not enabled, or sorting logic is not set correctly, resulting in results being returned only in the original recall order.
How to Verify Proper Configuration
- Embedding model configuration items are reviewed to confirm selected models match the platform-supported list, with no undefined model names present.
- A mix design notification document is uploaded, an index generation test is executed, and the chunk length is verified to match the preset value, with no errors during vector generation.
- A retrieval request for the aggregate moisture content correction logic is submitted, and the number of returned results is confirmed to align with the single search limit setting.
- The
defaultTokenconfiguration is checked to confirm it meets the average token consumption requirements of currently imported documents, preventing chunk truncation or index failure.
The values provided are common starting points and should be measured against the reader's own samples.