What the data for this category looks like
Daily data for this category includes packaging process cards, bonding and molding parameters, wafer incoming maps, test program versions, yield reports, failure analysis reports, and more. Data sources include wafer test maps, manufacturing execution systems (MES), bonding and molding equipment, ATE test machines, X-ray/SAM inspection equipment, and customer specification requirements. Data updates occur per production batch, per individual chip test, and once daily. Fields include wafer ID, test bin codes, process parameter values, failure analysis results, and others. Fields and units vary widely across different documents. For example, bonding parameters use degrees Celsius and megapascals as units, while test programs are identified by version numbers.
What constraints these characteristics impose on vector models and indexes
The data characteristics of this category create multiple constraints for the vector models and indexes link. First, multi-source heterogeneous data and high-frequency updates require indexes to support incremental updates and real-time synchronization. This prevents process parameter matching errors caused by delayed data. Second, diverse document types and large differences in field units require vector models to adapt to technical parameter text and structured chart descriptions. This ensures that embedded vectors can accurately distinguish between different test program versions and yield data. Third, a single production batch links to multiple documents. It is necessary to control the index's retrieval scope and calculation overhead, while ensuring matching result accuracy. This prevents serious issues such as mixed batches or incorrect program usage.
How to set configurations
These configurations are based on FastGPT v4.17.0, and are adapted for the vector models and indexes link in semiconductor packaging and testing scenarios:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the paragraph length of documents such as packaging process cards and test programs, avoiding truncation of key process parameters |
| Overlap Length | 100–200 characters | Retains parameter associations across chunks, preventing continuous parameters such as bonding temperature and molding pressure from being split |
| Similarity Threshold | 0.75–0.85 | Distinguishes subtle differences in technical parameters, avoiding matching to incorrect test program versions |
| Single Search Limit | Top 8 entries | Covers the number of associated documents for a single batch test, while controlling calculation overhead |
defaultToken | 4096 | Adapts to long-text content such as yield reports and failure analysis reports, covering the basic token consumption of a single query |
| Maximum Context | 16000 characters | Accommodates spliced content from multiple associated documents, supporting complete process specification queries |
Three common configuration mistakes
- A 401 error is returned when calling the vector model. Cause: The API key permissions for the vector model are not configured correctly, or the access parameters for the large language model and vector model are confused when binding the channel.
- Duplicate content is automatically deleted after custom-split documents are stored in the knowledge base, causing the index order to differ from the original document split results. Cause: The knowledge base duplicate document deduplication switch is not turned off, or chunk-level deduplication rules are not configured. The system uses default document-level deduplication, which causes order loss.
- The knowledge base index speed does not improve and may even decrease after deploying multiple K8s replicas. Cause: An index sharding strategy between replicas is not configured. Multiple replicas repeatedly pull the same data and create redundant indexes, increasing calculation overhead.
How to verify a correct configuration
- Upload a packaging process card document, and check whether the number of split chunks matches the expected chunk length range.
- Initiate a process parameter query, and check whether the number of returned matching results matches the single search limit setting.
- Test the vector model call, confirm that the returned embedding results have no errors and meet the required format.
- Simulate multi-batch data updates, and check whether the index update time meets business real-time requirements.
The values provided are common starting points and should be measured against the reader's own samples.