What Data This Category Includes
Daily data from injection molding, stamping, and die-casting mold enterprises, as well as mold centers of manufacturing enterprises. This data includes mold design task sheets, DFM reports, part drawings, processing process cards, trial mold reports, mold modification notices, and mold history records. Data sources cover customer product models, mold flow or forming simulation files, CAD/CAM output files, machining and measurement data, and records related to trial mold samples. Data is updated per project, with new records and documents generated after each trial mold or mold modification.
What Constraints These Characteristics Impose on Vector Models and Indexes
Data is updated iteratively per project and changes frequently. Each trial mold or mold modification adds or modifies documents. Vector indexes must support incremental updates to avoid time costs from full reconstruction. Documents include structured process parameters and defect records. Vector chunking must balance field integrity and semantic coherence, and avoid splitting key associated information such as mold numbers and trial mold dates. Traceability of mold quality problems requires accurate matching of associated content like defect types and mold modification history. Indexes must support multi-field associated recall, while controlling the amount of recalled data to ensure search speed.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Mold-related documents are mostly process descriptions and trial mold records. This single chunk length adapts to the input limit of most vector models while preserving semantic integrity |
| Overlap Length | 100–150 characters | Avoid semantic breaks caused by chunking, and ensure recall coherence of cross-chunk content such as trial mold defects and mold modification records |
| Similarity Threshold | 0.75–0.85 | Match the accuracy requirements for mold defect analysis, filter low-relevance historical records and reduce invalid recalls |
| Single Search Limit | Top 10 entries | Control the amount of vector recalled data to ensure search speed, while covering associated records for most typical quality problems |
maxToken | 4096 token | Adapt to the input limits of mainstream text embedding models, avoid indexing failures caused by token overrun |
defaultToken | Calibrated based on actual measurements | Adapt to the total token consumption of the current knowledge base, ensuring stability of batch indexing |
Three Common Misconfigurations
- Phenomenon: An error "No results available" is displayed when using the knowledge base. Cause: text-embedding-ada-002 is not correctly configured for the API channel, or the model is not associated with the currently used knowledge base.
- Phenomenon: The status "Indexing" is displayed indefinitely when accessing a local m3e vector model. Cause: Insufficient vector computing resources are configured, or the chunk length exceeds the processing limit of the model, causing the indexing process to block and fail to complete.
- Phenomenon: Search speed is too slow to meet real-time traceability requirements. Cause: The single search limit is not set, and the reranking link is not enabled, resulting in too much data recalled by vector search, increasing subsequent result processing time.
How to Verify Correct Configuration
- Enter the knowledge base settings page of FastGPT v4.17.0, check if the vector model channel has been added and enabled.
- Upload a mold trial report, perform manual indexing, and check if the indexing process is completed within 5 minutes.
- Enter a query about mold shrinkage defects, perform a search, and check if the returned results include relevant records of the corresponding trial mold report.
- Adjust the similarity threshold, and verify whether the matching degree of the recalled results meets the accuracy requirements of the current business.
The values provided are common starting points and should be measured against the reader's own samples.