What Data for This Category Looks Like
Photovoltaic module encapsulation process data comes from the cell grading system, glass adhesive film backplane batch ledgers, MES production system, string welding and lamination equipment logs, and EL/IV tester reports. Daily documents include layout drawings, material batch records, lamination process records, EL crack images, IV test reports, insulation pressure resistance records, and packing lists. Data updates in real time for each module and each material batch. Fields include batch number, process parameters, test values, equipment number, and more. Units include current gear, pressure, temperature, power, and other relevant units.
Constraints on Vector Models and Indexes
This category of data updates frequently, has a small single-data volume but a large total volume. This requires vector indexes to support high-concurrency writes and fast retrieval. Document types are diverse, including structured ledgers and unstructured charts. A vector embedding strategy must adapt to mixed-format content, avoiding embedding bias across different document types. Structured fields such as batch numbers and process parameters must be bound to unstructured text for indexing, ensuring accurate matching of material batches and corresponding process records during retrieval. High real-time requirements demand that index update latency be controlled within a reasonable range, preventing a disconnect between ledger data and retrieval results.
Configuration Settings
The configurations below are based on FastGPT v4.17.0.
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the average length of single documents such as lamination process records and IV test reports, avoiding embedding bias caused by overly long single chunks |
| Overlap Length | 100–150 characters | Retains batch association information across chunks, preventing chunks from breaking the binding between material batches and corresponding process records |
defaultToken | 4096 | Matches the input upper limit of locally deployed embedding models, ensuring that key fields such as material batch numbers and test values are fully embedded |
maxToken | 8192 | Adapts to the maximum context limit of the Qwen3-Embedding-8B model, avoiding truncation of long-text process parameters |
| Similarity Threshold | 0.75–0.85 | Matches the accuracy requirements for cell current gear matching, filtering low-correlation historical batch records |
| Single Search Limit | Top 20 entries | Covers the retrieval scope for multiple components per batch, ensuring that complete production data can be obtained during ledger summary |
Three Common Mistakes
- Milvus deployment fails with container startup failure, and logs show PostgreSQL connection timeout. The cause is failing to modify the Milvus compose file according to FastGPT official configurations. The default image includes redundant PostgreSQL dependencies, leading to excessive resource usage and failed startup.
- When connecting a locally deployed VLLM version of the Qwen3-Embedding-8B model, a model response timeout error occurs. The cause is failing to configure the correct model interface timeout in FastGPT; the default timeout setting is insufficient to cover the calculation delay of large model embedding.
- The storage location of vectorized documents cannot be located, making it impossible to verify data association relationships. The cause is failing to enable vector data path logging in FastGPT system settings, resulting in inability to query the binding relationship between storage directories and original documents.
How to Confirm Proper Configuration
- Upload a real lamination process record document, check that the chunked text fragments fully retain batch numbers and corresponding process parameters, with no critical content truncated.
- Initiate a retrieval for a specific cell current gear, verify that the matching degree of returned results falls within the preset similarity threshold range.
- View the FastGPT vector database storage directory, confirm that the association between embedded vector data and original documents is correct, with no missing fields.
- Test batch import of multi-batch documents, check that index updates are completed within the preset time range, with no accumulation errors.
The values provided here are common starting points and should be measured against the reader's own samples.