What Data for This Category Looks Like
The data for this category comes from multiple sources. These include BIM detailed design, rebar and concrete batch ledgers, mold table and curing kiln operation records, laboratory test reports, and component QR code traceability systems. Daily documents cover component detailed drawings, rebar cutting lists, mold inspection records, pouring and vibration logs, curing cycle records, and demolding and factory inspection reports. Updates occur per individual component, per batch of materials, and per curing kiln batch. All processes from production preparation to factory delivery are covered. Data fields include component number, material batch, curing duration, test strength, and other relevant fields. Units include millimeters, cubic meters, hours, megapascals, and additional standard units.
Constraints for Vector Models and Indexes
This category of data has high update frequency and fine granularity. Each individual component, batch of materials, and curing kiln batch generates new data. This creates clear requirements for the incremental update capability of indexes. Daily documents include long-text component detailed drawings and short-text test records. There are significant differences in data length. This requires adapting text chunking rules for different text lengths. Data is strongly bound to the QR code and batch number of individual components. Vector retrieval must associate traceability identifiers to accurately locate problem sources. Errors in the production link directly affect construction schedules and costs. Retrieval response speed must match business processing rhythms. This avoids affecting traceability efficiency due to slow index loading. FastGPT v4.17.0 supports vector model normalization configuration. This can adapt to non-normalized embedding models and ensure retrieval accuracy.
How to Set Configurations
| Configuration Item | Recommended Value/Range | Rationale |
|---|---|---|
| Maximum Context | 8000–12000 token | Matches the combined text length of precast concrete component detailed drawings and multiple associated records, to avoid truncating critical traceability information |
| Vector Model Normalization Configuration | Enabled | Adapts to non-normalized embedding models, ensuring accuracy of retrieval similarity calculations |
| Single Search Limit | Top 10 results | Balances retrieval recall rate and response speed, adapting to quality problem traceability requirements that link multiple documents |
defaultToken | 500 token | Serves as the default length for a single text chunk, adapting to short-text test records and rebar cutting list fragments |
maxToken | 1500 token | Limits the maximum chunk length, avoiding excessive index load caused by overly long component detailed drawing chunks |
| Similarity Threshold | 0.75–0.85 | Filters low-relevance retrieval results, focusing on production records directly related to quality problems |
Three Common Misconfigurations
- Symptom: After switching the knowledge base vector model, the background task status always shows "In Progress" with no progress. The original model cannot be switched back via the interface. Cause: Only the vector model configuration items were modified. Re-generation of the associated index was not triggered. This causes the system to fail to adapt to the vector format of the new model.
- Symptom: The retrieval results return a large number of irrelevant production records. Query duration exceeds 30 seconds. Cause: Reasonable single search limit and similarity threshold were not set. Long texts were not properly chunked. This leads to excessive index load.
- Symptom: A token limit exceeded error triggers during retrieval. The background shows token usage exceeds the preset limit. Cause:
maxTokenand maximum context parameters were not configured. Complete long-text component detailed drawings are directly sent to the retrieval process. This leads to excessive token consumption.
How to Verify Correct Configuration
- Navigate to the vector model configuration page. Confirm that the "Vector Model Normalization Configuration" switch is enabled.
- Submit a quality problem query. Verify that the number of returned results and similarity scores match the preset single search limit and similarity threshold.
- View the background task log. Confirm that the index reconstruction task starts and completes normally after switching the vector model.
- Test retrieval of a single short-text test record. Confirm that the
defaultTokenparameter takes effect and the chunk length meets expectations.
The values listed here are common starting points and should be evaluated against local samples.