What Data Looks Like for This Category
This category's data comes primarily from business systems and on-site collection at machining plants and precision parts factories. It includes documents such as process routes, operation sheets, tool lists, CNC programs, first-piece inspection records, patrol inspection forms, and measuring tool calibration records. Data sources cover product drawings and 3D models, CAPP/CAM, machine tool collection data, tool management systems, MES systems, and measuring and inspection tool data. Updates are triggered by work orders, single-piece production runs, sampling inspections, or tool changes. Most document fields include parameters with clear units, such as dimensional deviation, feed rate, and tool life.
Constraints Imposed on Vector Models and Indexing
The multi-source dispersion, high-frequency update frequency, and diverse document types of this category's data create multiple constraints for the vector models and indexing stage. Multi-source data includes both structured parameter documents and unstructured process text, so mixed-format chunking and vectorization rules must be supported. High-frequency update scenarios require indexes to support incremental synchronization, to avoid excessive delays from full index reconstruction. Documents contain a large number of process parameters with clear units. Vector models must retain the semantic association of these parameters, and avoid losing unit information. On-site collected raw data may have format inconsistencies, so basic format alignment must be completed before indexing.
Configuration Settings
This configuration applies to FastGPT v4.17.0.
| Configuration Item | Recommended Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the semantic integrity of single segments of documents such as CNC programs and process sheets, and avoids dispersing cross-operation information |
| Overlap Length | 100–200 characters | Retains association information between operations, and prevents key content such as benchmarks and clamping plans from being cut off by chunks |
| Similarity Threshold | 0.75–0.85 | Matches strongly correlated process parameters such as tool wear and dimensional deviation, and reduces false matching probability |
| Single Search Limit | Top 8 entries | Covers core associated data such as operations, tools, and inspection records involved in a single work order |
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 | Calibrate using representative process documents | This is the default token count used for direct text splitting; inspect sample chunks. It is separate from the embedding model input limit and LLM conversation context. See the v4.17.2 model-field reference |
Three Common Configuration Errors
- Phenomenon: Newly submitted tool change patrol inspection records do not appear in index search results. Cause: Incremental indexing is not configured for frequently updated documents such as tool lists and patrol inspection forms. Only full indexing is performed, so new data is not included in a timely manner.
- Phenomenon: Line numbers, feed rates and other parameters in retrieved CNC program fragments are broken. Cause: Chunk length is set too small, which cuts off continuous semantic units of CNC programs. This prevents vector retrieval from associating complete operation logic.
- Phenomenon: Custom retrieval processes cannot link operation sheets to first-piece inspection records. Cause: Entity association indexing configuration is not enabled. Only flat entity indexing is used, which cannot match cross-document operation-inspection association logic.
How to Verify Correct Configuration
- Upload a single CNC program and its corresponding operation sheet. Verify that retrieved associated documents include preset associated types such as tool lists and operation sheets, and confirm that the configured single search limit takes effect.
- Add a new tool change record. Verify that the index completes incremental synchronization without requiring a full indexing operation.
- Review vectorization process logs. Confirm that parameters with units such as dimensional deviation and feed rate in documents are fully included in the vector representation.
- Test retrieval results with different chunk lengths, and adjust to values that meet the semantic integrity requirements of this category's documents.
The values provided are common starting points and should be evaluated against the reader's own samples.