What Data for This Category Looks Like
Data sources for this category include resin batch property sheets, injection molding machine operating parameters, mold temperature controller records, MES system data, and dimensional and appearance inspection forms. Daily documents cover injection molding process cards, raw material drying records, mold history, first article inspection reports, patrol inspection forms, and material change and machine cleaning records. Data is updated every production batch, each mold or material change, and at the end of each shift. Document fields include resin grade, drying temperature, holding pressure, mold cavity temperature, detected dimensional tolerances, recycled material ratio, and other semi-structured process and production records.
Constraints Imposed on Vector Models and Indexes
Semi-structured process and production records require retention of field associations. Do not break the binding relationship between parameters and corresponding records during chunking. Frequently updated production data requires indexes to support incremental updates, avoiding the time cost of full reconstruction. Semantic differences in industrial fields such as resin grades and process parameters require more precise vector matching to avoid semantic misjudgments by general-purpose models. In bidding scenarios, historical process and quotation data must be associated. Indexes need to support associated retrieval of multiple document chunks, rather than only matching single text chunks.
Configuration Settings
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the length of single-segment parameter descriptions in injection molding process cards, avoiding cutting critical process parameters |
| Overlap Length | 100–150 characters | Retains front and back associated fields of material change and machine cleaning records, preventing loss of association between recycled material ratio and mold state after chunking |
| Similarity Threshold | 0.7–0.9 | Distinguishes property differences between different resin grades, avoiding mismatching of similar but different raw material parameters |
| Single Search Limit | Top 10 results | Matches the volume of historical process and quotation data required for bidding, balancing retrieval efficiency and recall completeness |
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 | 256 | Adapts to the basic embedding request length of single-segment injection molding process records |
| Maximum Context | 8000 tokens | Accommodates associated data from multiple injection molding process cards and quotation sheets, supporting parameter references in bid documents |
Three Common Mistakes
- Custom chunked document blocks arrive in incorrect order after import. The symptom is that the order of retrieved document chunks does not match the custom chunking order. The cause is that the system enables global deduplication by default, which deletes duplicate blocks while disrupting the order of custom chunked blocks.
- Vector model calls use the LLM channel. The symptom is that retrieval results do not meet semantic matching requirements, or error messages are returned. The cause is that the vector model was not bound to the knowledge base index link during custom channel configuration, and the LLM channel was mistakenly used as the vector entry.
- Index speed is too slow during bulk data import. The symptom is timeouts occur when importing multiple sets of injection molding machine parameter records. The cause is that incremental index updates were not configured, and full index reconstruction was executed for each import.
How to Verify Proper Configuration
- Import a single injection molding process card, check that the number of split chunks matches the custom chunk count, with no chunk loss caused by automatic deduplication.
- Submit a retrieval request for resin grades or process parameters, confirm that the returned similarity values fall within the 0–1 range and match the set threshold range.
- Import multiple sets of material change and machine cleaning records in bulk, check that index update time meets expectations, and no timeout errors occur.
- Confirm the vector model channel configuration under FastGPT v4.17.0, check that request parameters are not associated with the LLM channel.
All listed configuration values are common starting points, and should be validated against the reader's own sample data.