What the Data for This Category Looks Like
Data for this category comes from wafer test maps, MES systems, bonding and molding equipment logs, ATE tester outputs, X-ray/SAM inspection results, and customer specification documents. Daily documents include packaging process cards, bonding and molding parameter sheets, wafer incoming material maps, test program version files, yield statistics and failure analysis reports. Update frequency is per production batch, per chip test, and once daily. Document fields include wafer batch number, bonding temperature, molding pressure, test bin codes, equipment operating parameter values. Units include degrees Celsius, megapascals, volts, and others.
What Constraints These Characteristics Place on Vector Models and Indexes
Data sources for this category are scattered, including native device logs, system exported files, and drawing documents. Document formats vary widely across sources, so adaptation is required for splitting and vectorization processing of multiple text types. Data update frequency is high and granularity is fine. Document volume per production batch increases as production capacity rises, so the index must support high-frequency incremental updates. Documents contain a large number of process parameters with physical units and batch identifiers. Vector models must adapt to mixed content of structured numerical values and text. Content such as wafer incoming material maps and test bin codes includes strongly correlated batch information. The index must retain contextual associations to avoid batch misalignment during cross-review.
Configuration Settings
These configurations apply to FastGPT v4.17.0.
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the single-section content length of documents such as packaging process cards and parameter sheets, and avoids losing process parameter associations after splitting |
| Overlap Length | 50–100 characters | Retains the context of adjacent process steps, and prevents cross-block breakage of bonding and molding parameters |
| Similarity Threshold | 0.75–0.85 | Distinguishes minor differences in process parameters of the same batch, and avoids mismatching during mixed-batch review |
| Single Search Limit | Top 10 entries | Covers all document entries involved in a single batch production, and ensures comprehensiveness of cross-review |
defaultToken | 16384 | Adapts to the long-text content of failure analysis reports, and meets the vector generation requirements for complete context |
maxToken | 8192 | Limits the maximum number of tokens per segment during indexing, and balances vector generation quality and hardware resource usage |
Three Common Configuration Errors
- A 404 status code is returned after configuring baidu-embedding-v1. The cause is failure to correctly associate the OneAPI proxy address and model call permissions.
- Only one 3090 graphics card is used during index generation, and the second graphics card resource is not utilized. The cause is failure to enable multi-GPU scheduling logic in the deployment configuration.
- Imported packaging process cards lose adjacent bonding and molding parameters after splitting. The cause is failure to set a reasonable overlap length, causing cross-segment parameter association breakage.
How to Verify Correct Configuration
- A single packaging process card is imported. Split text blocks are reviewed. Process parameters are confirmed to not be split and broken. Chunk length and overlap length are adjusted to meet business requirements.
- Index generation logs are reviewed. Hardware resource usage is confirmed to match configured expectations. Multi-GPU scheduling is normal, or single-GPU load is reasonable.
- A single-batch cross-review request is submitted. The number of returned matching documents is verified to match the preset single search limit.
- Vector generation for long-text failure analysis reports is tested. No token limit exceeded error is confirmed.
maxTokenis adjusted to the applicable range.
The values provided are common starting points and should be evaluated against the reader’s own samples.