What Data for This Category Looks Like
Core 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 specifications. Daily generated documents cover packaging process cards, bonding and molding parameter tables, incoming wafer maps, test program versions, yield reports, and failure analysis reports. Document fields include batch identifiers, process parameters, die coordinates, test bin values, and more. Data update frequencies cover single-chip testing, each production batch, and daily summaries. Single-batch data updates in real time alongside production progress.
What Constraints These Characteristics Impose on Multi-turn Dialogue and Prompt Engineering
Dispersed data sources and frequent updates require that dialogue contexts be strictly bound to exclusive data for the current production batch, to avoid cross-batch confusion. Multi-turn dialogues must retain core information such as incoming wafer maps and test program versions for the current batch, to prevent retrieval errors caused by excessive context. Prompts must limit retrieval scope to process documents for the current work order, to avoid misuse of non-target documents such as failure analysis reports. It is also necessary to dynamically link the latest versions of equipment parameters, to ensure that bonding and molding parameters called during dialogues are valid configurations for the current batch.
How to Set Configurations
These configurations are based on FastGPT v4.17.0, adapted to process Q&A needs for semiconductor packaging and testing scenarios:
| Configuration Item | Recommended Setting | Rationale |
|---|---|---|
chatId | Bound to the unique identifier of the current production batch | Used to associate exclusive data such as incoming wafer maps and test program versions for this batch, to avoid cross-batch confusion |
variables | Pre-set to the current batch number and ATE model number | Used to dynamically insert retrieval scope into prompts, limiting calls to process documents for the corresponding batch |
messages | Retain no more than 8 rounds of context per single dialogue | Semiconductor process data updates frequently. Excessively long contexts will introduce data from old batches, reducing retrieval accuracy |
| Question optimization | Enabled | Process documents contain professional terms such as bonding pull force and molding reflow temperature. Optimization improves retrieval matching accuracy |
| Maximum context | Calibrated through actual testing, no more than 8000 tokens | A typical single process document is approximately 1000 characters long. 8000 tokens can hold 8 core documents, avoiding excessive context |
| Single search limit | Top 6 results | Covers core parameters for the current work order, avoiding redundant data interfering with dialogue results |
Three Common Mistakes
- Symptom: Knowledge base search works normally, but AI does not reference knowledge base content during dialogue. Cause: The retrieval scope for the current batch is not bound in
messages, so the dialogue context is not linked to preset knowledge base filter conditions. - Symptom: Program execution results cannot be used as background knowledge input to the next AI dialogue module. Cause: Execution results are not bound as dialogue variables through
variables, and the prompt does not reference this variable to pass content. - Symptom: Cross-batch process data appears in multi-turn dialogues, leading to incorrect parameter calls. Cause: The
chatIdbound to the current production batch is not set during initial dialogue, and context retains document data from previous batches.
How to Confirm Configurations Are Set Correctly
- Initiate a test dialogue, input "Current batch bonding parameters", check if retrieval results include batch documents bound to the current
chatId. - View the dialogue context panel, confirm that the batch number and ATE model number in
variableshave been correctly inserted into the system prompt. - Launch 3 consecutive dialogues for different work orders, check that the context only retains process data for the current batch, with no content from old batches.
- Trigger the program execution module to generate test results, check if subsequent dialogues reference this result as background knowledge.
The values provided here are common starting points, and should be evaluated against the reader's own samples.