What Data for This Category Looks Like
Daily documents for this category include packaging process cards, bonding and molding parameters, wafer incoming material maps, test program versions, yield reports, and failure analysis reports. Data sources cover wafer test maps, MES systems, bonding and molding equipment, ATE test machines, X-ray/SAM detection equipment, and customer specification requirements. Data is updated per production batch, per chip test node, and on a daily basis. Individual documents include fields such as equipment parameter thresholds, test bin values, wafer coordinate mapping, and others. Some drawing-type documents contain pixel-level coordinate information, with units including voltage, time, chip size parameters, and additional relevant metrics.
Constraints on Vector Models and Indexes From These Characteristics
Pixel-level coordinate information in wafer incoming material maps requires that chunking preserves associative integrity, avoiding splitting cross-coordinate mapping fields. Otherwise, the correspondence between wafer batches and physical objects will be lost. Format differences across multiple data sources require unified field mapping rules during index construction, to ensure consistent vector encoding logic for different device logs and test reports. High-frequency updated single-chip test data needs to support incremental indexing, to reduce the computational overhead of full index reconstruction. The length span of different document types is large, ranging from short texts such as test program versions to thousands-character failure analysis reports. This requires flexible chunking and encoding strategies.
Configuration Settings
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
| Chunk Length | 800–1200 characters | Adapts to the mixed text length of packaging process cards and failure analysis reports, avoiding splitting content with cross-parameter associations |
| Overlap Length | 50–80 characters | Preserves coordinate association information for wafer incoming material maps, preventing loss of critical mapping relationships at chunk boundaries |
defaultToken | Set from representative sample documents | defaultToken is the default token count for direct text chunking; tune it on representative source documents. See the v4.17.2 model-field reference. It is not a vector dimension, embedding-model input cap, or retrieval/LLM context. |
| Similarity Threshold | Calibrated via actual testing | Needs to be adjusted in combination with failure association features in yield reports, avoiding mismatched test data from different batches |
| Single Search Limit | Top 20 entries | Covers the single-batch test data volume of ATE test machines, ensuring sufficient recall of associated records |
maxToken | 4096 | Adapts to the long-text encoding needs of failure analysis reports, matching the maximum input limit of mainstream embedding models |
Three Common Misconfigurations
- Directly using the original knowledge base search after replacing the embedding model, leading to retrieval result deviations or no matching content. The cause is that the old index was generated based on the vector space of the original model, and the vector mapping rules were not updated synchronously.
- Setting
defaultTokenbeyond the input limit of the selected embedding model, resulting in model call errors, returning undefined model or parameter mismatch prompts. The cause is failure to align with the token limits of the model, and not adjusting parameters according to the selected model. - Sorting retrieval results only by similarity, without matching the priority logic of the business scenario, leading to low-association but high-similarity results interfering with audits. The cause is failure to configure a reranking link or bind business field weight rules.
How to Verify Proper Configuration
- Import a single packaging process card, check that the chunked text fragments retain complete bonding parameters and batch association information, with no splitting of cross-key fields.
- In FastGPT v4.17.0, trigger a full index rebuild after replacing the embedding model, verify that vector updates for all documents in the knowledge base are completed, with no entries in failed status.
- Input a test bin distribution query, check that the number of retrieved results matches the single search limit configuration, and the sorting logic conforms to the priority of business audits.
- Call the test interface, check that the
defaultTokenandmaxTokenparameters match the input requirements of the selected embedding model, with no parameter overlimit errors.
The values provided are common starting points and should be measured against the reader's own samples.