What Data for This Category Looks Like
Data sources for this category include wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM detection equipment, and customer specification documents. Daily documents include packaging process cards, bonding and molding parameter tables, incoming wafer maps, test program version files, yield reports, and failure analysis reports. Data update frequency follows each production batch, each individual chip test, and once per day. Document fields include process number, parameter threshold, equipment model, batch number, bin code, failure type, and more. Some documents include visual wafer dot maps and test timing data.
Constraints on the Vector Models and Indexes Workflow
The data includes visual dot maps and structured test parameters, resulting in large differences in unstructured text block lengths. Some test program files can reach tens of thousands of characters, so long text splitting rules must be adapted. Multi-source data has large differences in update frequency, so flexible switching between incremental indexing and full indexing must be supported. Embedding for wafer dot maps must match image vector models, and text vectors must use a unified index space. Yield reports from different batches have minor format differences, so adaptive text chunking logic must be configured to avoid chunk loss from format deviations. Exception reports involve highly specialized failure type terms. The vector model must ensure accurate encoding of professional terms to prevent semantic matching deviations.
Configuration Settings
| Configuration Item | Recommended Value | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the information density of single chunks for packaging process cards and yield reports, avoids chunk loss |
| Overlap Length | 100–150 characters | Retains professional term context across chunks, avoids semantic breaks |
| Similarity Threshold | 0.72–0.85 | Distinguishes matching accuracy between professional terms and common vocabulary, adapts to semantic differences of terms in the semiconductor industry |
| Single Recalled Entries | Top 8 entries | Covers the recall range of multi-source data, avoids missing key test parameters |
defaultToken | 1024 | Matches the token length of professional documents, ensures embedding completeness for long texts |
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. |
Three Common Misconfigurations
- The knowledge base remains in the indexing state for an extended period after upload. This occurs because multi-source data has large differences in update frequency, and the incremental indexing trigger logic does not align with the update rhythm of each batch and daily updates, leading to indexing queue backlogs.
- An error prompt about exceeding the embedding rate limit appears during vectorization. This is because the number of threads or model call quotas are not adjusted, and a reasonable chunk length is not configured, causing the token amount of a single call to exceed the limit.
- Chunk loss occurs when the chunk length is set to 3000 characters. This is because the paragraph structure of professional documents differs significantly from general texts. Excessively long chunks cause the model to fail to correctly identify semantic boundaries, resulting in content truncation and loss.
How to Verify Correct Configuration
- Upload a sample packaging process card, check that the number of split text chunks matches expectations, with no obvious content loss.
- Input a professional failure analysis question, confirm that the recalled knowledge base entries include the corresponding test program version and yield data.
- View the vector model call logs, verify that token usage does not exceed the
maxTokensetting range. - Test the incremental indexing function, upload a new batch yield report, confirm that the indexing status completes updating within a reasonable time frame.
The values provided here are common starting points, and should be evaluated against the reader's own samples.