What data looks like for this category
The data for this category mainly comes from silicon wafer incoming materials, MES systems, diffusion furnaces, coating equipment, printing presses, EL and IV testing equipment, and laboratory data. Update cycles cover each production batch, per-wafer testing, and daily summaries. Daily documents include process cards, diffusion and coating parameter tables, printing and sintering records, EL detection images, IV grading data, and yield reports. Fields include equipment operating temperature, coating thickness, crack coordinate positions, electrical performance parameters, batch numbers, and more. Units vary by data type, with no unified fixed standard.
Constraints on vector models and indexes
This category includes a mix of high-frequency real-time detection data (such as per-wafer EL images, IV data) and batch-level summary documents (such as process cards, yield reports). The index system must support fast retrieval of both small single files and large batch data. Multi-source heterogeneous data has significant field differences, so vector modeling must support structured and unstructured content in different formats. Daily batch-updated data creates index refresh pressure, so an incremental index mechanism must be configured to avoid excessive time spent on full reconstruction. Equipment inspection and maintenance Q&A needs to be associated with specific equipment parameters and historical fault records, so vector recall results must accurately match equipment numbers and production batches to avoid mixing data across batches.
Configuration Settings
The following configurations are based on FastGPT v4.17.0.
| Configuration Item | Recommended Value Range | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to common document lengths for this category, such as process parameter tables and EL detection image annotation text, to avoid breaking information within chunks |
| Overlap Length | 100–200 characters | Reduces loss of cross-chunk information, and adapts to continuous parameter records in batch summary documents |
| Similarity Threshold | 0.75–0.85 | Balances recall accuracy and recall volume, and matches the precise matching requirements of equipment inspection questions |
| Single Search Limit | Top 10 results | Controls context length to avoid exceeding model input limits |
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 token | Used for vector generation of short equipment inspection questions, balancing recall efficiency and semantic completeness |
Three Common Configuration Mistakes
- Symptom: When deploying a PG database via Docker, the knowledge base index fails to be created. The host is an 8c16G virtual machine without a GPU. Cause: The CPU acceleration mode for the vector plugin was not configured, leading to index calculation timeouts or excessive memory usage.
- Symptom: After importing Excel files and multi-dimensional documents from a team collaboration platform folder, the index does not include the document content. Cause: Parsing adaptation for the corresponding file formats was not enabled, and dedicated field mapping rules for photovoltaic category data were not configured.
- Symptom: Recall results for equipment inspection Q&A lack key process parameters. Cause: The chunk length was set too short, splitting continuous process parameter records into multiple chunks, leading to scattered key information.
How to Verify Correct Configuration
- Upload a typical process card document, check that the chunked text retains continuous parameter records completely, with no obvious breaks.
- Submit an equipment inspection question, verify that the recalled results' equipment numbers and production batches match the problem description, and adjust the similarity threshold to meet business requirements.
- View the index refresh log to confirm that daily batch-updated data only triggers incremental indexing, not full reconstruction.
- Test external API calls, confirm that the response time of the vector generation and retrieval process meets business expectations, and adjust the
maxTokenanddefaultTokenconfigurations to adapt to model limits.
The values provided are common starting points and should be measured against the reader's own samples.